https://ojs.aaai.org/index.php/AAAI-SS/issue/feed 2024-05-21T00:00:00-07:00 Publications Manager publications@aaai.org Open Journal Systems <p>The AAAI Symposium Series, previously published as AAAI Technical Reports, are held three times a year (Spring, Summer, Fall) and are designed to bring colleagues together to share ideas and learn from each other’s artificial intelligence research. The series affords participants a smaller, more intimate setting where they can share ideas and learn from each other’s artificial intelligence research. Topics for the symposia change each year, and the limited seating capacity and relaxed atmosphere allow for workshop-like interaction. The format of the series allows participants to devote considerably more time to feedback and discussion than typical one-day workshops. It is an ideal venue for bringing together new communities in emerging fields.<br /><br />The AAAI Spring Symposium Series is typically held during spring break (generally in March) on the west coast. The AAAI Summer Symposium Series is the newest in the annual set of meetings run in parallel at a common site. The inaugural 2023 Summer Symposium Series was held July 17-19, 2023, in Singapore. The AAAI Fall Symposium series is usually held on the east coast during late October or early November.</p> https://ojs.aaai.org/index.php/AAAI-SS/article/view/31170 2024-05-20T15:04:58-07:00 Cecilia O. Alm coagla@rit.edu

AI systems are breaking into new domains and applications, and it is pivotal to center humans in contemporary AI systems and contemplate what this means. This discussion considers three perspectives or human roles in AI as users, contributors, and researchers-in-training, to illustrate this notion.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31171 2024-05-20T15:04:59-07:00 Olivier Bartheye olivier.bartheye@ecole-air.fr Laurent Chaudron a2e357aad6ddbda7a5f2a2bf45e0ac3a@example.org

This present work is deliberately placed in the context capable of defining the requirements expressed by machine decision-making calculations. The informational nature of a decision requires abandoning any invariant preserving the structure but on the contrary switching into total chaos, a necessary and sufficient condition for exploiting the symmetries allowing the calculation to converge. Decision arithmetic is the best way to precisely define the nature of these symmetries.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31172 2024-05-20T15:05:00-07:00 Ellen J. Bass ellen.j.bass@drexel.edu Steven Weber a012c35989d55f3327ae2112a4330299@example.org

Current risk frameworks such as probabilistic risk analy-sis methodologies do not take societal safety-related benefits into account. To inform human-AI collaborative system development, this manuscript highlights the need for updated risk frameworks and suggestions for relevant considerations.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31173 2024-05-20T15:05:01-07:00 Beth Cardier bethcardier@hotmail.com)

How can a machine warn its human collaborator about an unexpected risk if the machine does not possess the explicit language required to name it? This research transfers techniques from creative writing into a conversational format that could enable a machine to convey a novel, open-world threat. Professional writers specialize in communicating unexpected conditions with inadequate language, using overlapping contextual and analogical inferences to adjust a reader’s situation model. This paper explores how a similar approach could be used in conversation by a machine to adapt its human collaborator’s situation model to include unexpected information. This method is necessarily bi-directional, as the process of refining unexpected meaning requires each side to check in with each other and incrementally adjust. A proposed method and example is presented, set five years hence, to envisage a new kind of capability in human-machine interaction. A near-term goal is to develop foundations for autonomous communication that can adapt across heterogeneous contexts, especially when a trusted outcome is critical. A larger goal is to make visible the level of communication above explicit communication, where language is collaboratively adapted.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31174 2024-05-20T15:05:02-07:00 Wanyi Chen wanyi.chen503@duke.edu Mary Cummings ca2ac938126c88725482c7f20d9f797c@example.org

Model selection is a necessary step in unsupervised machine learning. Despite numerous criteria and metrics, model selection remains subjective. A high degree of subjectivity may lead to questions about repeatability and reproducibility of various machine learning studies and doubts about the robustness of models deployed in the real world. Yet, the impact of modelers' preferences on model selection outcomes remains largely unexplored. This study uses the Hidden Markov Model as an example to investigate the subjectivity involved in model selection. We asked 33 participants and three Large Language Models (LLMs) to make model selections in three scenarios. Results revealed variability and inconsistencies in both the participants’ and the LLMs' choices, especially when different criteria and metrics disagree. Sources of subjectivity include varying opinions on the importance of different criteria and metrics, differing views on how parsimonious a model should be, and how the size of a dataset should influence model selection. The results underscore the importance of developing a more standardized way to document subjective choices made in model selection processes.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31175 2024-05-20T15:05:03-07:00 Yaliang Chuang yaliang.chuang@gmail.com Poyang David Huang p.huang@student.tue.nl

Aesthetics is a crucial aspect of design that plays a critical role in the creation process and customers' perception of outcomes. However, aesthetic expressions are highly subjective and nuanced. It often relies on designers' experiences and many trials and errors to get it right. Our research first investigated how designers and artists curated aesthetic materials and utilized them in their daily practice. Based on the result, we applied Langley's human-like learning framework to develop an interactive Style Agent system. It aims to learn designers' aesthetic expertise and utilize AI's capability to empower practitioner's creativity. In this paper, we used typographic posters as examples and conducted a preliminary evaluation of our prototype. The results showed that our system provided a modular structure for effortlessly annotating users' subjective perceptions and making the visualizations easy to interpret through performance. Overall, it acts as a facilitator to help enhance their own aesthetic awareness and empowers them to expand their design space.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31176 2024-05-20T15:05:05-07:00 Boris Galitsky bgalitsky@hotmail.com

In spite of the great progress of large language models (LLMs) in recent years, there is a popular belief that their limitations need to be addressed “from outside”, by building hybrid neurosymbolic systems which add robustness, explainability, perplexity and verification done at a symbolic level. We propose shape-charged learning in the form of Meta-learning/DNN - kNN that enables the above features by integrating LMM with explainable nearest neighbor learning (kNN) to form the object-level, having deductive reasoning-based metalevel control learning processes, performing validation and correction of predictions in a way that is more interpretable by humans.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31177 2024-05-20T15:05:06-07:00 Ted Goranson tedgoranson@icloud.com

We explore a novel approach to complex domain modelling by emphasising primitives based on perception. The usual approach either focuses on actors or cognition associated with tokens that convey information. In related research, we have examined using effects and/or outcomes as primitives, and influences as the generator of those outcomes via categoric functors. That approach (influences, effects) has advantages: it leverages what is known and supports the expanded logics we use, where we want to anticipate and engineer possible futures. But it has weaknesses when placed in a dynamic human-machine system where what is perceived or assumed matters more than what is known. The work reported here builds on previous advances in type specification and reasoning to ‘move the primitives forward’ more toward situation encounter and away from situation understanding. The goal is in the context of shared human-machine systems where: • reaction times are shorter than the traditional ingestion/comprehension/response loop can support; • situations that are too complex or dynamic for current comprehension by any means; • there simply is insufficient knowledge about governing situations for the comprehension model to support action; and/or, • the many machine/human and system/system interfaces that are incapable of conveying the needed insights; that is, the communication channels choke the information or influence flows. While the approach is motivated by the above unfriendly conditions, we expect significant benefits. We will explore these but engineer toward a federated decision paradigm where decisions by local human, machine or synthesis are not whole-situation-aware, but that collectively ‘swarm’ locally across the larger system to be more effective, ‘wiser’ than a convention paradigm may produce. The supposed implementation strategy will be through extending an existing ‘playbooks as code’ project whose goals are to advise on local action by modelling and gaming complex system dynamics. A sponsoring context is ‘grey zone’ competition that avoids armed conflict, but that can segue to a mixed system course of action advisory. The general context is a costly ‘blue swan’ risk in large commercial and government enterprises. The method will focus on patterns and relationships in synthetic categories used to model type transitions within topological models of system influence. One may say this is applied intuitionistic type theory, following mechanisms generally described by synthetic differential geometry. In this context, the motivating supposition of this study is that information-carrying influence channels are best modelled in our challenging domain as perceived types rather than understood types.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31178 2024-05-20T15:05:08-07:00 Shih-Hong Huang szh277@psu.edu Ting-Hao 'Kenneth' Huang 188f46a2a21bc258c21325cac8a8819f@example.org

It is easy to assume that Large Language Models (LLMs) will seamlessly take over applications, especially those that are largely automated. In the case of conversational voice assistants, commercial systems have been widely deployed and used over the past decade. However, are we indeed on the cusp of the future we envisioned? There exists a social-technical gap between what people want to accomplish and the actual capability of technology. In this paper, we present a case study comparing two voice assistants built on Amazon Alexa: one employing a human-in-the-loop workflow, the other utilizes LLM to engage in conversations with users. In our comparison, we discovered that the issues arising in current human-in-the-loop and LLM systems are not identical. However, the presence of a set of similar issues in both systems leads us to believe that focusing on the interaction between users and systems is crucial, perhaps even more so than focusing solely on the underlying technology itself. Merely enhancing the performance of the workers or the models may not adequately address these issues. This observation prompts our research question: What are the overlooked contributing factors in the effort to improve the capabilities of voice assistants, which might not have been emphasized in prior research?

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31179 2024-05-20T15:05:09-07:00 Simon Kapiamba simon.t.kapiamba.civ@us.navy.mil Hesham Fouad 87ecd773f3631f8e98a62a002dadc042@example.org Ira S. Moskowitz 9580eda746178774c392c7c45c8e5012@example.org

This paper outlines an approach for assessing and quantifying the risks associated with integrating Large Language Models (LLMs) in generating naval operational plans. It aims to explore the potential benefits and challenges of LLMs in this context and to suggest a methodology for a comprehensive risk assessment framework.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31180 2024-05-20T15:05:10-07:00 Thuy Ngoc Nguyen ngoc.nguyen@udayton.edu Chase McDonald chasemcd@andrew.cmu.edu Cleotilde Gonzalez coty@cmu.edu

Temporal credit assignment is the process of distributing delayed outcomes to each action in a sequence, which is essential for learning to adapt and make decisions in dynamic environments. While computational methods in reinforcement learning, such as temporal difference (TD), have shown success in tackling this issue, it remains unclear whether these mechanisms accurately reflect how humans handle feedback delays. Furthermore, cognitive science research has not fully explored the credit assignment problem in humans and cognitive models. Our study uses a cognitive model based on Instance-Based Learning Theory (IBLT) to investigate various credit assignment mechanisms, including equal credit, exponential credit, and TD credit, using the IBL decision mechanism in a goal-seeking navigation task with feedback delays and varying levels of decision complexity. We compare the performance and process measures of the different models with human decision-making in two experiments. Our findings indicate that the human learning process cannot be fully explained by any of the mechanisms. We also observe that decision complexity affects human behavior but not model behavior. By examining the similarities and differences between human and model behavior, we summarize the challenges and opportunities for developing learning agents that emulate human decisions in dynamic environments.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31181 2024-05-20T15:05:11-07:00 Stephen Russell stephen.russell@jhsmiami.org Fabio Montes Suros noemail01@example.com Ashwin Kumar noemail02@example.com

For a large healthcare system, ignoring costs associated with managing the patient encounter denial process (staffing, contracts, etc.), total denial-related amounts can be more than $1B annually in gross charges. Being able to predict a denial before it occurs has the potential for tremendous savings. Using machine learning to predict denial has the potential to allow denial-preventing interventions. However, challenges of data imbalance make creating a single generalized model difficult. We employ two biased models in a hybrid voting scheme to achieve results that exceed the state-of-the art and allow for incremental predictions as the encounter progresses. The model had the added benefit of monitoring the human-driven denial process that affect the underlying distribution, on which the models’ bias is based.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31182 2024-05-20T15:05:13-07:00 Geoffrey W. Rutledge geoff@healthtap.com Alexander Sivura 6c6eccc4924ddad6cc034da060adc386@example.org

We describe "Dr. A.I.", a virtual physician assistant that uses generative AI to conduct a pre-visit patient interview and to create a draft clinical note for the physician. We document the effectiveness of Dr. A.I. by measuring the concordance of the actual diagnosis made by the doctor with the generated differ-ential diagnosis (DDx) list. This application demonstrates the practical healthcare capabilities of a large language model to improve efficiency of doctor visits while also addressing safety concerns for the use of generative AI in the workflow of patient care.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31183 2024-05-20T15:05:13-07:00 Diyi Yang diyiy@stanford.edu

Large language models (LLMs) have revolutionized the way humans interact with AI systems, transforming a wide range of fields and disciplines. In this talk, I share two distinct approaches to empowering human-AI interaction using LLMs. The first one explores how LLMstransform computational social science, and how human-AI collaboration can reduce costs and improve the efficiency of social science research. The second part looks at social skill learning via LLMs by empowering therapists and learners with LLM-empowered feedback and deliberative practices. These two works demonstrate how human-AI collaboration via LLMs can empower individuals and foster positive change. We conclude by discussing how LLMs enable collaborative intelligence by redefining the interactions between humans and AI systems.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31184 2024-05-20T15:05:15-07:00 Ming Yin mingyin@purdue.edu

Artificial intelligence (AI) technologies have been increasingly integrated into human workflows. For example, the usage of AI-based decision aids in human decision-making processes has resulted in a new paradigm of AI-assisted decision making---that is, the AI-based decision aid provides a decision recommendation to the human decision makers, while humans make the final decision. The increasing prevalence of human-AI collaborative decision making highlights the need to understand how humans engage with the AI-based decision aid in these decision-making processes, and how to promote the effectiveness of the human-AI team in decision making. In this talk, I'll discuss a few examples illustrating that when AI is used to assist humans---both an individual decision maker or a group of decision makers---in decision making, people's engagement with the AI assistance is largely subject to their heuristics and biases, rather than careful deliberation of the respective strengths and limitations of AI and themselves. I'll then describe how to enhance AI-assisted decision making by accounting for human engagement behavior in the designs of AI-based decision aids. For example, AI recommendations can be presented to decision makers in a way that promotes their appropriate trust and reliance on AI by leveraging or mitigating human biases, informed by the analysis of human competence in decision making. Alternatively, AI-assisted decision making can be improved by developing AI models that can anticipate and adapt to the engagement behavior of human decision makers.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31185 2024-05-20T15:05:19-07:00 Nils Beutling nils.beutling@students.fhnw.ch Maja Spahic-Bogdanovic maja.spahic@fhnw.ch

This paper presents a Knowledge-Based Recommender System (KBRS) that aims to align course recommendations with students' career goals in the field of information systems. The developed KBRS uses the European Skills, Competences, qualifications, and Occupations (ESCO) ontology, course descriptions, and a Large Language Model (LLM) such as ChatGPT 3.5 to bridge course content with the skills required for specific careers in information systems. In this context, no reference is made to the previous behavior of students. The system links course content to the skills required for different careers, adapts to students' changing interests, and provides clear reasoning for the courses proposed. An LLM is used to extract learning objectives from course descriptions and to map the promoted competency. The system evaluates the degree of relevance of courses based on the number of job-related skills supported by the learning objectives. This recommendation is supported by information that facilitates decision-making. The paper describes the system's development, methodology and evaluation and highlights its flexibility, user orientation and adaptability. It also discusses the challenges that arose during the development and evaluation of the system.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31186 2024-05-20T15:05:21-07:00 Stefan Dernbach stefan.dernbach@pnnl.gov Khushbu Agarwal khushbu.agarwal@pnnl.gov Alejandro Zuniga alejandro.michelzuniga@pnnl.gov Michael Henry michael.j.henry@pnnl.gov Sutanay Choudhury sutanay.choudhury@pnnl.gov

Integrating large language models with knowledge graphs derived from domain-specific data represents an important advancement towards more powerful and factual reasoning. As these models grow more capable, it is crucial to enable them to perform multi-step inferences over real-world knowledge graphs while minimizing hallucination. While large language models excel at conversation and text generation, their ability to reason over domain-specialized graphs of interconnected entities remains limited. For example, can we query a model to identify the optimal contact in a professional network for a specific goal, based on relationships and attributes in a private database? The answer is no – such capabilities lie beyond current methods. However, this question underscores a critical technical gap that must be addressed. Many high-value applications in areas such as science, security, and e-commerce rely on proprietary knowledge graphs encoding unique structures, relationships, and logical constraints. We introduce a fine-tuning framework for developing Graph-aligned Language Models (GaLM) that transforms a knowledge graph into an alternate text representation with labeled question-answer pairs. We demonstrate that grounding the models in specific graph-based knowledge expands the models’ capacity for structure-based reasoning. Our methodology leverages the large-language model's generative capabilities to create the dataset and proposes an efficient alternate to retrieval-augmented generation styled methods.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31187 2024-05-20T15:05:22-07:00 Charles Dickens cadicken@ucsc.edu Connor Pryor cfpryor@ucsc.edu Lise Getoor getoor@ucsc.edu

Neural-symbolic (NeSy) AI strives to empower machine learning and large language models with fast, reliable predictions that exhibit commonsense and trustworthy reasoning by seamlessly integrating neural and symbolic methods. With such a broad scope, several taxonomies have been proposed to categorize this integration, emphasizing knowledge representation, reasoning algorithms, and applications. We introduce a knowledge representation-agnostic taxonomy focusing on the neural-symbolic interface capturing methods that reason with probability, logic, and arithmetic constraints. Moreover, we derive expressions for gradients of a prominent class of learning losses and a formalization of reasoning and learning. Through a rigorous empirical analysis spanning three tasks, we show NeSy approaches reach up to a 37% improvement over neural baselines in a semi-supervised setting and a 19% improvement over GPT-4 on question-answering.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31188 2024-05-20T15:05:24-07:00 Florian Geissler florian.geissler@iks.fraunhofer.de Karsten Roscher karsten.roscher@iks.fraunhofer.de Mario Trapp mario.trapp@iks.fraunhofer.de

Generative AI is increasingly important in software engineering, including safety engineering, where its use ensures that software does not cause harm to people. This also leads to high quality requirements for generative AI. Therefore, the simplistic use of Large Language Models (LLMs) alone will not meet these quality demands. It is crucial to develop more advanced and sophisticated approaches that can effectively address the complexities and safety concerns of software systems. Ultimately, humans must understand and take responsibility for the suggestions provided by generative AI to ensure system safety. To this end, we present an efficient, hybrid strategy to leverage LLMs for safety analysis and Human-AI codesign. In particular, we develop a customized LLM agent that uses elements of prompt engineering, heuristic reasoning, and retrieval-augmented generation to solve tasks associated with predefined safety concepts, in interaction with a system model graph. The reasoning is guided by a cascade of micro-decisions that help preserve structured information. We further suggest a graph verbalization which acts as an intermediate representation of the system model to facilitate LLM-graph interactions. Selected pairs of prompts and responses relevant for safety analytics illustrate our method for the use case of a simplified automated driving system.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31189 2024-05-20T15:05:25-07:00 Sadaf Ghaffari sadafgh@colostate.edu Nikhil Krishnaswamy nkrishn87@gmail.com

In this paper, we present an exploration of LLMs' abilities to problem solve with physical reasoning in situated environments. We construct a simple simulated environment and demonstrate examples of where, in a zero-shot setting, both text and multimodal LLMs display atomic world knowledge about various objects but fail to compose this knowledge in correct solutions for an object manipulation and placement task. We also use BLIP, a vision-language model trained with more sophisticated cross-modal attention, to identify cases relevant to object physical properties that that model fails to ground. Finally, we present a procedure for discovering the relevant properties of objects in the environment and propose a method to distill this knowledge back into the LLM.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31190 2024-05-20T15:05:27-07:00 Filippos Gouidis gouidis@ics.forth.gr Katerina Papantoniou papanton@ics.forth.gr Konstantinos Papoutsakis kpapoutsakis@hmu.gr Theodore Patkos patkos@csd.uoc.gr Antonis Argyros argyros@ics.forh.gr Dimitris Plexousakis dp@ics.forth.gr

Domain-specific knowledge can significantly contribute to addressing a wide variety of vision tasks. However, the generation of such knowledge entails considerable human labor and time costs. This study investigates the potential of Large Language Models (LLMs) in generating and providing domain-specific information through semantic embeddings. To achieve this, an LLM is integrated into a pipeline that utilizes Knowledge Graphs and pre-trained semantic vectors in the context of the Vision-based Zero-shot Object State Classification task. We thoroughly examine the behavior of the LLM through an extensive ablation study. Our findings reveal that the integration of LLM-based embeddings, in combination with general-purpose pre-trained embeddings, leads to substantial performance improvements. Drawing insights from this ablation study, we conduct a comparative analysis against competing models, thereby highlighting the state-of-the-art performance achieved by the proposed approach.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31191 2024-05-20T15:05:29-07:00 Nicholas Harvel nicholas.harvel@moduleq.com Felipe Bivort Haiek felipe.bivort@moduleq.com Anupriya Ankolekar anupriya@moduleq.com David James Brunner djb@moduleq.com

Large Language Models (LLMs) can increase the productivity of general-purpose knowledge work, but accuracy is a concern, especially in professional settings requiring domain-specific knowledge and reasoning. To evaluate the suitability of LLMs for such work, we developed a benchmark of 16 analytical tasks representative of the investment banking industry. We evaluated LLM performance without special prompting, with relevant information provided in the prompt, and as part of a system giving the LLM access to domain-tuned functions for information retrieval and planning. Without access to functions, state-of-the-art LLMs performed poorly, completing two or fewer tasks correctly. Access to appropriate domain-tuned functions yielded dramatically better results, although performance was highly sensitive to the design of the functions and the structure of the information they returned. The most effective designs yielded correct answers on 12 out of 16 tasks. Our results suggest that domain-specific functions and information structures, by empowering LLMs with relevant domain knowledge and enabling them to reason in domain-appropriate ways, may be a powerful means of adapting LLMs for use in demanding professional settings.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31192 2024-05-20T15:05:31-07:00 Jia Huang jia.huang@tamu.edu Peng Jiang maskjp@tamu.edu Alvika Gautam alvikag@tamu.edu Srikanth Saripalli ssaripalli@tamu.edu

Predicting pedestrian behavior is the key to ensure safety and reliability of autonomous vehicles. While deep learning methods have been promising by learning from annotated video frame sequences, they often fail to fully grasp the dynamic interactions between pedestrians and traffic, crucial for accurate predictions. These models also lack nuanced common sense reasoning. Moreover, the manual annotation of datasets for these models is expensive and challenging to adapt to new situations. The advent of Vision Language Models (VLMs) introduces promising alternatives to these issues, thanks to their advanced visual and causal reasoning skills. To our knowledge, this research is the first to conduct both quantitative and qualitative evaluations of VLMs in the context of pedestrian behavior prediction for autonomous driving. We evaluate GPT-4V(ision) on publicly available pedestrian datasets: JAAD and WiDEVIEW. Our quantitative analysis focuses on GPT-4V's ability to predict pedestrian behavior in current and future frames. The model achieves a 57% accuracy in a zero-shot manner, which, while impressive, is still behind the state-of-the-art domain-specific models (70%) in predicting pedestrian crossing actions. Qualitatively, GPT-4V shows an impressive ability to process and interpret complex traffic scenarios, differentiate between various pedestrian behaviors, and detect and analyze groups. However, it faces challenges, such as difficulty in detecting smaller pedestrians and assessing the relative motion between pedestrians and the ego vehicle.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31193 2024-05-20T15:05:33-07:00 Milan Kostic milan.kostic@unicam.it Hans Friedrich Witschel hansfriedrich.witschel@fhnw.ch Knut Hinkelmann knut.hinkelmann@fhnw.ch Maja Spahic-Bogdanovic maja.spahic@fhnw.ch

This study delves into the application of large language models (LLMs), such as ChatGPT-4, for the automated evaluation of student essays, with a focus on a case study conducted at the Swiss Institute of Business Administration. It explores the effectiveness of LLMs in assessing German-language student transfer assignments, and contrasts their performance with traditional evaluations by human lecturers. The primary findings highlight the challenges faced by LLMs in terms of accurately grading complex texts according to predefined categories and providing detailed feedback. This research illuminates the gap between the capabilities of LLMs and the nuanced requirements of student essay evaluation. The conclusion emphasizes the necessity for ongoing research and development in the area of LLM technology to improve the accuracy, reliability, and consistency of automated essay assessments in educational contexts.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31194 2024-05-20T15:05:34-07:00 Emanuele Laurenzi emanuele.laurenzi@fhnw.ch Adrian Mathys adrian.mathys@students.fhnw.ch Andreas Martin andreas.martin@fhnw.ch

Conventional knowledge engineering approaches aiming to create Enterprise Knowledge Graphs (EKG) still require a high level of manual effort and high ontology expertise, which hinder their adoption across industries. To tackle this issue, we explored the use of Large Language Models (LLMs) for the creation of EKGs through the lens of a design-science approach. Findings from the literature and from expert interviews led to the creation of the proposed artefact, which takes the form of a six-step process for EKG development. Scenarios on how to use LLMs are proposed and implemented for each of the six steps. The process is then evaluated with an anonymised data set from a large Swiss company. Results demonstrate that LLMs can support the creation of EKGs, offering themselves as a new aid for knowledge engineers.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31195 2024-05-20T15:05:35-07:00 Pei-Ying Lin pearlie.lin10@gmail.com Erick Chandra erickchandra.1@gmail.com Jane Yung-jen Hsu yjhsu@csie.ntu.edu.tw

Commonsense reasoning refers to the ability to make inferences, draw conclusions, and understand the world based on general knowledge and commonsense. Whether Large Language Models (LLMs) have commonsense reasoning ability remains a topic of debate among researchers and experts. When confronted with multiple-choice commonsense reasoning tasks, humans typically rely on their prior knowledge and commonsense to formulate a preliminary answer in mind. Subsequently, they compare this preliminary answer to the provided choices, and select the most likely choice as the final answer. We introduce Aggregated Semantic Matching Retrieval (ASMR) as a solution for multiple-choice commonsense reasoning tasks. To mimic the process of humans solving commonsense reasoning tasks with multiple choices, we leverage the capabilities of LLMs to first generate the preliminary possible answers through open-ended question which aids in enhancing the process of retrieving relevant answers to the question from the given choices. Our experiments demonstrate the effectiveness of ASMR on popular commonsense reasoning benchmark datasets, including CSQA, SIQA, and ARC (Easy and Challenge). ASMR achieves state-of-the-art (SOTA) performance with a peak of +15.3% accuracy improvement over the previous SOTA on SIQA dataset.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31196 2024-05-20T15:05:37-07:00 Carsten Maletzki carsten.maletzki@dfki.de Eric Rietzke c6cb103429600468781ad2ffb0060a9c@example.org Ralph Bergmann 128c8249c077c0340da9f927bc46c5cd@example.org

Hybrid intelligence systems aim to leverage synergies in closely collaborating teams of humans and artificial intelligence (AI). To guide the realization of such teams, recent research proposed design patterns that capture role-based knowledge on human-AI collaborations. Building on these patterns requires hybrid intelligence systems to provide mechanisms that orchestrate human and AI contributions accordingly. So far, it is unclear if such mechanisms can be provided based on shared representations of the required knowledge. In this regard, we expect ontology-based data-centric process modeling to be a promising direction for hybrid intelligence systems that aim to support knowledge-intensive processes (KiPs). We illustrate this through exemplary process models (realized with our ontology- and data-driven business process model -- ODD-BP) that reflect the team design patterns for hybrid intelligence systems. We point out that relying on such process models enables multiple actors to fulfill roles jointly and allows them to address individual shortcomings. This is examined by discussing integrating large language models (LLMs) into the process models and describing how complementary AI actors could help to empower LLMs to fulfill their role in human-AI collaboration more comprehensively. Future work will extend the provided concepts while their evaluation initially focuses on the KiP of medical emergency call handling.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31197 2024-05-20T15:05:38-07:00 Andreas Martin andreas.martin@fhnw.ch Charuta Pande charuta.pande@fhnw.ch Sandro Schwander sandro.schwander@fhnw.ch Ademola J. Ajuwon ajajuwon@gmail.com Christoph Pimmer christoph.pimmer@swisstph.ch

FAQs are widely used to respond to users’ knowledge needs within knowledge domains. While LLM might be a promising way to address user questions, they are still prone to hallucinations i.e., inaccurate or wrong responses, which, can, inter alia, lead to massive problems, including, but not limited to, ethical issues. As a part of the healthcare coach chatbot for young Nigerian HIV clients, the need to meet their information needs through FAQs is one of the main coaching requirements. In this paper, we explore if domain knowledge in HIV FAQs can be represented as text embeddings to retrieve similar questions matching user queries, thus improving the understanding of the chatbot and the satisfaction of the users. Specifically, we describe our approach to developing an FAQ chatbot for the domain of HIV. We used a predefined FAQ question-answer knowledge base in English and Pidgin co-created by HIV clients and experts from Nigeria and Switzerland. The results of the post-engagement survey show that the chatbot mostly understood the user’s questions and could identify relevant matching questions and retrieve an appropriate response.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31198 2024-05-20T15:05:40-07:00 Andreas Martin andreas.martin@fhnw.ch Charuta Pande charuta.pande@fhnw.ch Hans Friedrich Witschel hansfriedrich.witschel@fhnw.ch Judith Mathez judith.mathez@fhnw.ch

We propose conversational agents as a means to simulate expert interviews, integrated into a simulational learning environment: ChEdventure. Designing and developing conversational agents using the existing tools and frameworks requires technical knowledge and a considerable learning curve. Recently, LLMs are being leveraged for their adaptability to different domains and their ability to perform various tasks in a natural, human-like conversational style. In this work, we explore if LLMs can help educators easily create conversational agents for their individual teaching goals. We propose a generalized template-based approach using LLMs that can instantiate conversational agents as an integrable component of teaching and learning activities. We evaluate our approach using prototypes generated from this template and identify guidelines to improve the experience of educators.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31199 2024-05-20T15:05:42-07:00 Andreas Martin andreas.martin@fhnw.ch Hans Friedrich Witschel hansfriedrich.witschel@fhnw.ch Maximilian Mandl maximilian.mandl@nagra.ch Mona Stockhecke mona.stockhecke@nagra.ch

This position paper presents a novel approach of semantic verification in Large Language Model-based Retrieval Augmented Generation (LLM-RAG) systems, focusing on the critical need for factually accurate information dissemination during public debates, especially prior to plebiscites e.g. in direct democracies, particularly in the context of Switzerland. Recognizing the unique challenges posed by the current generation of Large Language Models (LLMs) in maintaining factual integrity, this research proposes an innovative solution that integrates retrieval mechanisms with enhanced semantic verification processes. The paper outlines a comprehensive methodology following a Design Science Research approach, which includes defining user personas, designing conversational interfaces, and iteratively developing a hybrid dialogue system. Central to this system is a robust semantic verification framework that leverages a knowledge graph for fact-checking and validation, ensuring the correctness and consistency of information generated by LLMs. The paper discusses the significance of this research in the context of Swiss direct democracy, where informed decision-making is pivotal. By improving the accuracy and reliability of information provided to the public, the proposed system aims to support the democratic process, enabling citizens to make well-informed decisions on complex issues. The research contributes to advancing the field of natural language processing and information retrieval, demonstrating the potential of AI and LLMs in enhancing civic engagement and democratic participation.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31200 2024-05-20T15:05:43-07:00 Orfeas Menis Mastromichalakis menisorfeas@gmail.com Edmund Dervakos eddiedervakos@islab.ntua.gr Alexandros Chortaras achort@cs.ntua.gr Giorgos Stamou gstam@softlab.ntua.gr

The use of symbolic knowledge representation and reasoning as a way to resolve the lack of transparency of machine learning classifiers is a research area that has lately gained a lot of traction. In this work, we use knowledge graphs as the underlying framework providing the terminology for representing explanations for the operation of a machine learning classifier escaping the constraints of using the features of raw data as a means to express the explanations, providing a promising solution to the problem of the understandability of explanations. In particular, given a description of the application domain of the classifier in the form of a knowledge graph, we introduce a novel theoretical framework for representing explanations of its operation, in the form of query-based rules expressed in the terminology of the knowledge graph. This allows for explaining opaque black-box classifiers, using terminology and information that is independent of the features of the classifier and its domain of application, leading to more understandable explanations but also allowing the creation of different levels of explanations according to the final end-user.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31201 2024-05-20T15:05:45-07:00 Ankur Padia pankur1@umbc.edu Francis Ferraro ferraro@umbc.edu Tim Finin finin@cs.umbc.edu

High-quality knowledge graphs (KGs) play a crucial role in many applications. However, KGs created by automated information extraction systems can suffer from erroneous extractions or be inconsistent with provenance/source text. It is important to identify and correct such problems. In this paper, we study leveraging the emergent reasoning capabilities of large language models (LLMs) to detect inconsistencies between extracted facts and their provenance. With a focus on ``open'' LLMs that can be run and trained locally, we find that few-shot approaches can yield an absolute performance gain of 2.5-3.4% over the state-of-the-art method with only 9% of training data. We examine the LLM architectures' effect and show that Decoder-Only models underperform Encoder-Decoder approaches. We also explore how model size impacts performance and counterintuitively find that larger models do not result in consistent performance gains. Our detailed analyses suggest that while LLMs can improve KG consistency, the different LLM models learn different aspects of KG consistency and are sensitive to the number of entities involved.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31202 2024-05-20T15:05:46-07:00 Kristina Radivojevic kradivo2@nd.edu Nicholas Clark nclark3@nd.edu Paul Brenner paul.r.brenner@nd.edu

The emergence of Large Language Models (LLMs) has great potential to reshape the landscape of many social media platforms. While this can bring promising opportunities, it also raises many threats, such as biases and privacy concerns, and may contribute to the spread of propaganda by malicious actors. We developed the "LLMs Among Us" experimental framework on top of the Mastodon social media platform for bot and human participants to communicate without knowing the ratio or nature of bot and human participants. We built 10 personas with three different LLMs, GPT-4, Llama 2 Chat, and Claude. We conducted three rounds of the experiment and surveyed participants after each round to measure the ability of LLMs to pose as human participants without human detection. We found that participants correctly identified the nature of other users in the experiment only 42% of the time despite knowing the presence of both bots and humans. We also found that the choice of persona had substantially more impact on human perception than the choice of mainstream LLMs.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31203 2024-05-20T15:05:47-07:00 Kanak Raj imh10032.19@bitmesra.ac.in Kaushik Roy kaushikr@email.sc.edu Vamshi Bonagiri vamshib2@umbc.edu Priyanshul Govil pgovil1@umbc.edu Krishnaprasad Thirunarayan t.k.prasad@wright.edu Raxit Goswami raxit.g@shaip.com Manas Gaur manas@umbc.edu

Personalizing conversational agents can enhance the quality of conversations and increase user engagement. However, they often lack external knowledge to appropriately tend to a user’s persona. This is crucial for practical applications like mental health support, nutrition planning, culturally sensitive conversations, or reducing toxic behavior in conversational agents. To enhance the relevance and comprehensiveness of personalized responses, we propose using a two-step approach that involves (1) selectively integrating user personas and (2) contextualizing the response by supplementing information from a background knowledge source. We develop K-PERM (Knowledge-guided PErsonalization with Reward Modulation), a dynamic conversational agent that combines these elements. K-PERM achieves state-of-the- art performance on the popular FoCus dataset, containing real-world personalized conversations concerning global landmarks.We show that using responses from K-PERM can improve performance in state-of-the-art LLMs (GPT 3.5) by 10.5%, highlighting the impact of K-PERM for personalizing chatbots.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31204 2024-05-20T15:05:49-07:00 Kaushik Roy kaushikr@email.sc.edu Alessandro Oltramari alessandro.oltramari@us.bosch.com Yuxin Zi yzi@email.sc.edu Chathurangi Shyalika jayakodc@email.sc.edu Vignesh Narayanan vignar@sc.edu Amit Sheth amit@sc.edu

Large Language Models have excelled at encoding and leveraging language patterns in large text-based corpora for various tasks, including spatiotemporal event-based question answering (QA). However, due to encoding a text-based projection of the world, they have also been shown to lack a full bodied understanding of such events, e.g., a sense of intuitive physics, and cause-and-effect relationships among events. In this work, we propose using causal event graphs (CEGs) to enhance language understanding of spatiotemporal events in language models, using a novel approach that also provides proofs for the model’s capture of the CEGs. A CEG consists of events denoted by nodes, and edges that denote cause and effect relationships among the events. We perform experimentation and evaluation of our approach for benchmark spatiotemporal QA tasks and show effective performance, both quantitative and qualitative, over state-of-the-art baseline methods.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31205 2024-05-20T15:05:50-07:00 Sagar Srinivas Sakhinana sagar.sakhinana@tcs.com Geethan Sannidhi geethansannidhi20@cse.iiitp.ac.in Venkataramana Runkana venkat.runkana@tcs.com

We present a novel framework for analyzing and interpreting electron microscopy images in semiconductor manufacturing using vision-language instruction tuning. The framework employs a unique teacher-student approach, leveraging pretrained multimodal large language models such as GPT-4 to generate instruction-following data for zero-shot visual question answering (VQA) and classification tasks, customizing smaller multimodal models (SMMs) for microscopy image analysis, resulting in an instruction tuned language-and-vision assistant. Our framework merges knowledge engineering with machine learning to integrate domain-specific expertise from larger to smaller multimodal models within this specialized field, greatly reducing the need for extensive human labeling. Our study presents a secure, cost-effective, and customizable approach for analyzing microscopy images, addressing the challenges of adopting proprietary models in semiconductor manufacturing.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31206 2024-05-20T15:05:52-07:00 Jaelle Scheuerman jaelle.scheuerman@nrlssc.navy.mil Dina Acklin dina.acklin@nrlssc.navy.mil

Many behavioral science studies result in large amounts of unstructured data sets that are costly to code and analyze, requiring multiple reviewers to agree on systematically chosen concepts and themes to categorize responses. Large language models (LLMs) have potential to support this work, demonstrating capabilities for categorizing, summarizing, and otherwise organizing unstructured data. In this paper, we consider that although LLMs have the potential to save time and resources performing coding on qualitative data, the implications for behavioral science research are not yet well understood. Model bias and inaccuracies, reliability, and lack of domain knowledge all necessitate continued human guidance. New methods and interfaces must be developed to enable behavioral science researchers to efficiently and systematically categorize unstructured data together with LLMs. We propose a framework for incorporating human feedback into an annotation workflow, leveraging interactive machine learning to provide oversight while improving a language model's predictions over time.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31207 2024-05-20T15:05:53-07:00 Ben Schuering ben.schuering@gmail.com Thomas Schmid schmid@informatik.uni-leipzig.de

In recent years, artificial intelligence (AI) has seen significant advances that have in fact exceeded even optimistic prognoses. Using data-driven AI, namely deep learning techniques, it has been demonstrated that computers may now be equipped with abilities of remarkable scope and quality, such as solving image and text processing tasks at human level. Large language models, in particular, have sparked debates regarding opportunities and challenges of this rapidly developing area. Will remaining fundamental challenges of data-driven AI, such as factual or logical mistakes, be overcome for good if complemented and hybridized with symbolic AI techniques, such as knowledge representation and reasoning? Will systems of artificial general intelligence (AGI) emerge from this, possessing common sense and in fact completing the decades-old quest for AI that motivated the raise of the field in the 1950s? In the light of these questions, we review the likewise, decades-old philosophical debate about capabilities and limitations of computers from a hybrid AI point of view. Here, we discuss how hybrid AI is coming closer to disproving Hubert Dreyfus’ famous statements regarding what computers can not do. At the same time, we shed light on a lesser discussed challenge for hybrid AI: the possibility that its developers might be its biggest limiters.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31208 2024-05-20T15:05:54-07:00 Lucas L. Snijder l.l.snijder@student.tue.nl Quirine T. S. Smit quirine.smit@tno.nl Maaike H. T. de Boer maaike.deboer@tno.nl

One of the approaches to help the demand and supply problem in the labor market domain is to change from degree-based hiring to skill-based hiring. The link between occupations, degrees and skills is captured in domain ontologies such as ESCO in Europe and O*NET in the US. Several countries are also building or extending these ontologies. The alignment of the ontologies is important, as it should be clear how they all relate. Aligning two ontologies by creating a mapping between them is a tedious task to do manually, and with the rise of generative large language models like GPT-4, we explore how language models and domain knowledge can be combined in the matching of the instances in the ontologies and in finding the specific relation between the instances (mapping refinement). We specifically focus on the process of updating a mapping, but the methods could also be used to create a first-time mapping. We compare the performance of several state-of-the-art methods such as GPT-4 and fine-tuned BERT models on the mapping between ESCO and O*NET and ESCO and CompetentNL (the Dutch variant) for both ontology matching and mapping refinement. Our findings indicate that: 1) Match-BERT-GPT, an integration of BERT and GPT, performs best in ontology matching, while 2) TaSeR outperforms GPT-4, albeit marginally, in the task of mapping refinement. These results show that domain knowledge is still important in ontology alignment, especially in the updating of a mapping in our use cases in the labor domain.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31209 2024-05-20T15:05:56-07:00 Neşet Özkan Tan tan.neset@gmail.com Niket Tandon nikett@allenai.org David Wadden davidw@allenai.org Oyvind Tafjord oyvindt@allenai.org Mark Gahegan m.gahegan@auckland.ac.nz Michael Witbrock m.witbrock@auckland.ac.nz

Claim verification in scientific domains requires models that faithfully incorporate relevant knowledge from the ever-growing, vast existing literature. Unfaithful claim verifications can lead to misinformation such as those observed during the COVID-19 pandemic. Fact-checking systems often fail to capture the complex relationship between claims and evidence, especially with ambiguous claims and implicit assumptions. Relying only on current LLMs poses challenges due to hallucinations and information traceability issues. To address these challenges, our approach considers multiple viewpoints onto the scientific literature, enabling the assessment of contradictory arguments and implicit assumptions. Our proposed inference method adds faithful reasoning to large language models by distilling information from diverse, relevant scientific abstracts. This method provides a verdict label that can be weighted by the reputation of the scientific articles and an explanation that can be traced back to sources. Our findings demonstrate that humans not only perceive our explanation to be significantly superior to the off-the-shelf model, but they also evaluate it as faithfully enabling the tracing of evidence back to its original sources.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31210 2024-05-20T15:05:57-07:00 Christopher Toukmaji christoukmaji@gmail.com Allison Tee ateecup@stanford.edu

We present BIDARA, a Bio-Inspired Design And Research Assistant, to address the complexity of biomimicry -- the practice of designing modern-day engineering solutions inspired by biological phenomena. Large Language Models (LLMs) have been shown to act as sufficient general-purpose task solvers, but they often hallucinate and fail in regimes that require domain-specific and up-to-date knowledge. We integrate Retrieval-Augmented Generation (RAG) and Reasoning-and-Action agents to aid LLMs in avoiding hallucination and utilizing updated knowledge during generation of biomimetic design solutions. We find that incorporating RAG increases the feasibility of the design solutions in both prompting and agent settings, and we use these findings to guide our ongoing work. To the extent of our knowledge, this is the first work that integrates and evaluates Retrieval-Augmented Generation within LLM-generated biomimetic design solutions.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31211 2024-05-20T15:05:59-07:00 Yuxin Zi yzi@email.sc.edu Kaushik Roy kaushikr@email.sc.edu Vignesh Narayanan vignar@sc.edu Amit Sheth amit@sc.edu

Despite their extensive application in language understanding tasks, large language models (LLMs) still encounter challenges including hallucinations - occasional fabrication of information - and alignment issues - lack of associations with human-curated world models (e.g., intuitive physics or common-sense knowledge). Moreover, the black-box nature of LLMs presents significant obstacles in training them effectively to achieve desired behaviors. In particular, modifying the concept embedding spaces of LLMs can be highly intractable. This process involves analyzing the implicit impact of such adjustments on the myriad parameters within LLMs and the resulting inductive biases. We propose a novel architecture that wraps powerful function approximation architectures within an outer, interpretable read-out layer. This read-out layer can be scrutinized to explicitly observe the effects of concept modeling during the training of the LLM. Our method stands in contrast with gradient-based implicit mechanisms, which depend solely on adjustments to the LLM parameters and thus evade scrutiny. By conducting extensive experiments across both generative and discriminative language modeling tasks, we evaluate the capabilities of our proposed architecture relative to state-of-the-art LLMs of similar sizes. Additionally, we offer a qualitative examination of the interpretable read-out layer and visualize the concepts it captures. The results demonstrate the potential of our approach for effectively controlling LLM hallucinations and enhancing the alignment with human expectations.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31212 2024-05-20T15:06:03-07:00 Sree Bhargavi Balija sbalija@ucsd.edu Amitash Nanda ananda@ucsd.edu Debashis Sahoo dsahoo@ucsd.edu

Large language models (LLM) have gathered attention with the advent of ChatGPT. However, developing personalized LLM models faces challenges in real-world applications due to data scarcity and privacy concerns. Federated learning addresses these issues, providing collaborative training while preserving the client’s data. Although it has made significant progress, federated learning still faces ongoing challenges, such as communication efficiency, heterogeneous data, and privacy-preserving methods. This paper presents a novel, fully decentralized federated learning framework for LLMs to address these challenges. We utilize different blockchain-federated LLM (BC-FL) algorithms, effectively balancing the trade-off between latency and accuracy in a decentralized-federated learning environment. Additionally, we address the challenge of communication overhead in peer-to-peer networks by optimizing the path for weight transfer and mitigating node anomalies. We conducted experiments to evaluate memory usage and latency in server and serverless environments. Our results demonstrate a decrease in latency by 5X and a 13% increase in accuracy for serverless cases. Comparisons between synchronous and asynchronous scenarios revealed a 76% reduction in information passing time for the latter. The PageRank method is most efficient in eliminating anomalous nodes for better performance of the global federated LLM model. The code is available on GitHub (https://github.com/Sreebhargavibalijaa/Federated_finetuning_LLM-s_p2p_environment)

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31213 2024-05-20T15:06:05-07:00 Nathalie Baracaldo baracald@us.ibm.com

Federated learning (FL) has arisen as an alternative to collecting large amounts of data in a central place to train a machine learning (ML) model. FL is privacy-friendly, allowing multiple parties to collaboratively train an ML model without exchanging or transmitting their training data. For this purpose, an aggregator iteratively coordinates the training process among parties, and parties simply share with the aggregator model updates, which contain information pertinent to the model such as neural network weights. Besides privacy, generalization has been another key driver for FL: parties who do not have enough data to train a good performing model by themselves can now engage in FL to obtain an ML model suitable for their tasks. Products and real applications in the industry and consumer space have demonstrated the power of this learning paradigm. Recently, foundation models have taken the AI community by storm, promising to solve the shortage of labeled data. A foundation model is a powerful model that can be recycled for a variety of use cases by applying techniques such as zero-shot learning and full or parameter-efficient fine tuning. The premise is that the amount of data required to fine tune a foundation model for a new task is much smaller than fully training a traditional model from scratch. The reason why this is the case is that a good foundation model has already learned relevant general representations, and thus, adapting it to a new task only requires a minimal number of additional samples. This raises the question: Is FL still alive in the era of foundation models? In this talk, I will address this question. I will present some use cases where FL is very much alive. In these use cases, finding a foundation model with a desired representation is difficult if not impossible. With this pragmatic point of view, I hope to shed some light into a real use case where disparate private data is available in isolation at different parties and where labels may be located at a single party that doesn’t have any other information, making it impossible for a single party to train a model on its own. Furthermore, in some vertically-partitioned scenarios, cleaning data is not an option due to privacy-related reasons and it is not clear how to apply foundation models. Finally, I will also go over a few other requirements that are often overlooked, such as unlearning of data and its implications for the lifecycle management of FL and systems based on foundation models.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31214 2024-05-20T15:06:07-07:00 Yiran Chen yiran.chen@duke.edu

In the emerging field of federated learning (FL), the challenge of heterogeneity, both in data and systems, presents significant obstacles to efficient and effective model training. This talk focuses on the latest advancements and solutions addressing these challenges. The first part of the talk delves into data heterogeneity, a core issue in FL, where data distributions across different clients vary widely and affect FL convergence. We will introduce the FedCor framework addressing this by modeling loss correlations between clients using Gaussian Process and reducing expected global loss. External covariate shift in FL is uncovered, demonstrating that normalization layers are crucial, and layer normalization proves effective. Additionally, class imbalance in FL degrades performance, but our proposed Federated Class-balanced Sampling (Fed-CBS) mechanism reduces this imbalance by employing homomorphic encryption for privacy preservation. The second part of the talk shifts focus to system heterogeneity, an equally critical challenge in FL. System heterogeneity involves the varying computational capabilities, network speeds, and other resource-related constraints of participating devices in FL. To address this, we introduce FedSEA, which is a semi-asynchronous FL framework that addresses accuracy drops by balancing aggregation frequency and predicting local update arrival. Additionally, we discuss FedRepre, a framework specifically designed to enhance FL in real-world environments by addressing challenges including unbalanced local dataset distributions, uneven computational capabilities, and fluctuating network speeds. By introducing a client selection mechanism and a specialized server architecture, FedRepre notably improves the efficiency, scalability, and performance of FL systems. Our talk aims to provide a comprehensive overview of the current research and advancements in tackling both data and system heterogeneity in federated learning. We hope to highlight the path forward for FL, underlining its potential in diverse real-world applications while maintaining data privacy and optimizing resource usage.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31215 2024-05-20T15:06:08-07:00 Mark J. Gerken mark.gerken@baesystems.us

You can’t get more “on the tactical edge” than in space. No other operational domain suffers from the combinations of distance from the operator, harsh environments, unreachable assets with aging hardware, and increadably long communications as space systems. The complexity of developing and deploying AI solutions in satellites and probes is far more difficult than deploying similar AI on Earth. This talk explores some of the considerations involved in deploying AI and machine learning (ML) in the space domain.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31216 2024-05-20T15:06:09-07:00 Seyyedali Hosseinalipour alipour@buffalo.edu

In this work, we aim to unveil a new class of intermediate FL architectures between centralized and decentralized schemes called “FedCrawl.” FedCrawl takes advantage of benefits of D2D communications similar to decentralized schemes; however, it uses them in a nuanced way. FedCrawl is inspired by web crawlers, which effectively explore the websites to find updated/new content posted on the internet. The cornerstone of FedCrawl is its innovative conceptualization of neural networks (NNs) or other used ML models as autonomous entities, called random walkers, with the capability to move or jump across nodes in the network through peer-to-peer (P2P) or device-to-device (D2D) connections. We introduce five research aspects to study the nuanced intricacies governing random walker behavior in these environments. The first research aspect addresses the interplay between network topology and data distribution, emphasizing the importance of considering both factors for designing efficient random walks in FedCrawl. The second research aspect explores the applicability of node importance metrics in optimizing random walker paths for FedCrawl. We propose a dynamic perception-aware design, discussed in the third research aspect, where transition matrices adapt to the evolving state of random walkers, balancing exploration and exploitation. The fourth research aspect introduces innovative features like skipping, memory look-back, and caching/trailing to enhance random walker performance. Lastly, the fifth research aspect delves into the dynamics of multiple random walkers in networked environments, introducing the concept of multi-pole random walkers. Complementing these five research aspects, we present five conjectures, each introducing novel perspectives and methodologies in the domain of decentralized learning. These conjectures encompass areas such as temperature-based characterization of random walkers and network nodes, dynamic transition matrices, non-Markovian processes, and an evolutionary framework for random walker patterns.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31217 2024-05-20T15:06:11-07:00 Anurata Prabha Hridi aphridi@ncsu.edu Rajeev Sahay r2sahay@ucsd.edu Seyyedali Hosseinalipour alipour@buffalo.edu Bita Akram bakram@ncsu.edu

The majority of current research on the application of artificial intelligence (AI) and machine learning (ML) in science, technology, engineering, and mathematics (STEM) education relies on centralized model training architectures. Typically, this involves pooling data at a centralized location alongside an ML model training module, such as a cloud server. However, this approach necessitates transferring student data across the network, leading to privacy concerns. In this paper, we explore the application of federated learning (FL), a highly recognized distributed ML technique, within the educational ecosystem. We highlight the potential benefits FL offers to students, classrooms, and institutions. Also, we identify a range of technical, logistical, and ethical challenges that impede the sustainable implementation of FL in the education sector. Finally, we discuss a series of open research directions, focusing on nuanced aspects of FL implementation in educational contexts. These directions aim to explore and address the complexities of applying FL in varied educational settings, ensuring its deployment is technologically sound, beneficial, and equitable for all stakeholders involved.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31218 2024-05-20T15:06:12-07:00 Jithin Jagannath jjagannath@androcs.com

Machine learning promises to empower dynamic resource allocation requirements of Next Generation (NextG) wireless networks including 6G and tactical networks. Recently, we have seen the impact machine learning can make on various aspects of wireless networks. Yet, in most cases, the progress has been limited to simulations and/or relies on large processing units to run the decision engines as opposed to deploying it on the radio at the edge. While relying on simulations for rapid and efficient training of deep reinforcement learning (DRL) may be necessary, it is key to mitigate the sim-real gap while trying to improve the generalization capability. To mitigate these challenges, we developed the Marconi-Rosenblatt Framework for Intelligent Networks (MR-iNet Gym), an open-source architecture designed for accelerating the deployment of novel DRL for NextG wireless networks. To demonstrate its impact, we tackled the problem of distributed frequency and power allocation while emphasizing the generalization capability of DRL decision engine. The end-to-end solution was implemented on the GPU-embedded software-defined radio and validated using over-the-air evaluation. To the best of our knowledge, these were the first instances that established the feasibility of deploying DRL for optimized distributed resource allocation for next-generation of GPU-embedded radios.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31219 2024-05-20T15:06:14-07:00 Hana Khamfroush khamfroush@uky.edu

In a resource constrained environment like Internet-of-Things (IoT) systems, it is critical to make optimal decisions on how much resources to allocate pre-processing and how much to allocate to model training, and which specific combination of preprocessing and learning should be selected. This talk first, provides an overview of some initial steps we took towards developing federated data pre-processing in IoT environments, and then a visionary overview of potential research problems related to developing an integrated resource-aware and Quality-of-Service (QoS)-aware data pre-processing and model training system is provided.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31220 2024-05-20T15:06:15-07:00 Sanmi Koyejo sanmi@cs.stanford.edu

Machine learning (ML) models are routinely trained and deployed among distributed devices, e.g., learning with geo-distributed data centers and federated learning with mobile devices. Such shared computing platforms are susceptible to hardware, software, communication errors, and security concerns. This talk will outline some of the threat models in distributed learning, along with robust learning methods proposed to augment the fault tolerance of distributed machine learning, showing both theoretical and empirical evidence of robustness to benign and adversarial attacks.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31221 2024-05-20T15:06:16-07:00 Scott Kuzdeba scott.kuzdeba@baesystems.us

The Internet of Things (IoT) has revolutionized how our devices are networked, connecting multiple aspects of our life from smart homes and wearables to smart cities and warehouses. IoT’s strength comes from the ever-expanding diverse heterogeneous sensors, applications, and concepts that are all centered around the core concept collecting and sharing data from sensors. Simultaneously, deep learning has changed how our systems operate, allowing them to learn from data and change the way we interface with the world. Federated learning moves these two paradigm shifts together, leveraging the data (securely) from the IoT to train deep learning architectures for performant edge applications. However, today’s federated learning has not yet benefited from the scale of diversity that the IoT and deep learning sensors and applications provide. This talk explores how we can better tap into the heterogeneity that surrounds the potential of federated learning and use it to build better models. This includes the heterogeneity from device hardware to training paradigms (supervised, unsupervised, reinforcement, self-supervised).

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31222 2024-05-20T15:06:17-07:00 Aritra Mitra amitra2@ncsu.edu

Stochastic gradient descent (SGD) is at the heart of large-scale distributed machine learning paradigms such as federated learning (FL). In these applications, the task of training high-dimensional weight vectors is distributed among several workers that exchange information over networks of limited bandwidth. While parallelization at such an immense scale helps to reduce the computational burden, it creates several other challenges: delays, asynchrony, and most importantly, a significant communication bottleneck. The popularity and success of SGD can be attributed in no small part to the fact that it is extremely robust to such deviations from ideal operating conditions. Inspired by these findings, we ask: Are common reinforcement learning (RL) algorithms also robust to similarly structured perturbations? Perhaps surprisingly, despite the recent surge of interest in multi-agent/federated RL, almost nothing is known about the above question. This paper collects some of our recent results in filling this void.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31223 2024-05-20T15:06:18-07:00 Jose Angel Sanchez Viloria josesanchez2019@fau.edu Dimitris Stripelis stripeli@isi.edu Panos P. Markopoulos panagiotis.markopoulos@utsa.edu George Sklivanitis gsklivanitis@fau.edu Dimitris A. Pados dpados@fau.edu

The ability to rapidly understand and label the radio spectrum in an autonomous way is key for monitoring spectrum interference, spectrum utilization efficiency, protecting passive users, monitoring and enforcing compliance with regulations, detecting faulty radios, dynamic spectrum access, opportunistic mesh networking, and numerous NextG regulatory and defense applications. We consider the problem of automatic modulation classification (AMC) by a distributed network of wireless sensors that monitor the spectrum for signal transmissions of interest over a large deployment area. Each sensor receives signals under a specific channel condition depending on its location and trains an individual model of a deep neural network (DNN) accordingly to classify signals. To improve modulation classification accuracy, we consider federated learning (FL) where each individual sensor shares its trained model with a centralized controller, which, after aggregation, initializes its model for the next round of training. Without exchanging any spectrum data (such as in cooperative spectrum sensing), this process is repeated over time. A common DNN is built across the net- work while preserving the privacy associated with signals collected at different locations. Given their distributed nature, the statistics of the data across these sensors are likely to differ significantly. We propose the use of adaptive federated learning for AMC. Specifically, we use FEDADAM -an algorithm using Adam for server optimization – and ex- amine how it compares to the FEDAVG algorithm -one of the standard FL algorithms, which averages client parameters after some local iterations, in particular in challenging scenarios that include class imbalance and/or noise-level imbalance across the network. Our extensive numerical studies over 11 standard modulation classes corroborate the merit of adaptive FL, outperforming its standard alternatives in various challenging cases and for various network sizes.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31224 2024-05-20T15:06:19-07:00 Ashish Shenoy ashishvshenoy@gmail.com Sid Wang yuwang2020@meta.com Pierce Chuang pichuang@meta.com John Nguyen ngjhn@meta.com

In recent years, Federated Learning (FL) has shown significant advancements in its ability to perform various natural language processing (NLP) tasks. This work focuses on applying personalized FL for on-device language modeling. Due to limitations of memory and latency, these models cannot support the complexity of sub-word tokenization or beam search decoding, resulting in the decision to deploy a closed-vocabulary language model. However, closed-vocabulary models are unable to handle out-of-vocabulary (OOV) words belonging to specific users. To address this issue, We propose a novel technique called "OOV expansion" that improves OOV coverage and increases model accuracy while minimizing the impact on memory and latency. This method introduces a personalized "OOV adapter" that effectively transfers knowledge from a central model and learns word embedding for personalized vocabulary. OOV expansion significantly outperforms standard FL personalization methods on a set of common FL benchmarks.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31225 2024-05-20T15:06:21-07:00 Nojan Sheybani nsheyban@ucsd.edu Farinaz Koushanfar fkoushanfar@ucsd.edu

Inherently, federated learning (FL) robustness is very challenging to guarantee, especially when trying to maintain privacy. Compared to standard ML settings, FL's open training process allows for malicious clients to easily go under the radar. Alongside this, malicious clients can easily collude to attack the training process continuously, and without detection. FL models are also still susceptible to attacks on standard ML training procedures. This massive attack surface makes balancing the tradeoff between utility, practicality, robustness, and privacy extremely challenging. While there have been proposed defenses to attacks using popular privacy-preserving primitives, such as fully homomorphic encryption, they often face trouble balancing an all-important question that is present in all privacy-preserving systems: How much utility and practicality am I willing to give up to ensure privacy and robustness? In this work, we discuss a practical approach towards secure and robust FL and the challenges that face this field of emerging research.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31226 2024-05-20T15:06:22-07:00 Nicholas Soures nichsoures@noemail.com Vedant Karia Karia@noemail.com Dhireesha Kudithipudi dhireesha.kudithipudi@utsa.edu

Lifelong learning, which refers to an agent's ability to continuously learn and enhance its performance over its lifespan, is a significant challenge in artificial intelligence (AI), that biological systems tackle efficiently. This challenge is further exacerbated when AI is deployed in untethered environments with strict energy and latency constraints. We take inspiration from neural plasticity and investigate how to leverage and build energy-efficient lifelong learning machines. Specifically, we study how a combination of neural plasticity mechanisms, namely neuromodulation, synaptic consolidation, and metaplasticity, enhance the continual learning capabilities of AI models. We further co-design architectures that leverage compute-in-memory topologies and sparse spike-based communication with quantization for the edge. Aspects of this co-design can be transferred to federated lifelong learning scenarios.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31227 2024-05-20T15:06:23-07:00 Mehreen Tahir mehreen.tahir16@gmail.com Muhammad Intizar Ali ali.intizar@dcu.ie

Federated Learning (FL) has received tremendous attention as a decentralized machine learning (ML) framework that allows distributed data owners to collaboratively train a global model without sharing raw data. Since FL trains the model directly on edge devices, the heterogeneity of participating clients in terms of data distribution, hardware capabilities and network connectivity can significantly impact the overall performance of FL systems. Optimizing for model accuracy could extend the training time due to the diverse and resource-constrained nature of edge devices while minimizing training time could compromise the model's accuracy. Effective client selection thus becomes crucial to ensure that the training process is not only efficient but also capitalizes on the diverse data and computational capabilities of different devices. To this end, we propose FedPROM, a novel framework that tackles client selection in FL as a multi-criteria optimization problem. By leveraging the PROMETHEE method, FedPROM ranks clients based on their suitability for a given FL task, considering multiple criteria such as system resources, network conditions, and data quality. This approach allows FedPROM to dynamically select the most appropriate set of clients for each learning round, optimizing both model accuracy and training efficiency. Our evaluations on diverse datasets demonstrate that FedPROM outperforms several state-of-the-art FL client selection protocols in terms of convergence speed, and accuracy, highlighting the framework's effectiveness and the importance of multi-criteria client selection in FL.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31228 2024-05-20T15:06:25-07:00 Elahe Vedadi elahevedadi@google.com Joshua V. Dillon jvdillon@google.com Philip Andrew Mansfield memes@google.com Karan Singhal karansinghal@google.com Arash Afkanpour arashaf@google.com Warren Richard Morningstar wmorning@google.com

Conventional federated learning algorithms train a single global model by leveraging all participating clients’ data. However, due to heterogeneity in client generative distributions and predictive models, these approaches may not appropriately approximate the predictive process, converge to an optimal state, or generalize to new clients. We study personalization and generalization in stateless cross-device federated learning setups assuming heterogeneity in client data distributions and predictive models. We first propose a hierarchical generative model and formalize it using Bayesian Inference. We then approximate this process using Variational Inference to train our model efficiently. We call this algorithm Federated Variational Inference (FedVI). We use PAC-Bayes analysis to provide generalization bounds for FedVI. We evaluate our model on FEMNIST and CIFAR-100 image classification and show that FedVI beats the state-of-the-art on both tasks.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31229 2024-05-20T15:06:26-07:00 Lun Wang lunwang@google.com

In this talk, we will discuss how to make federated learning secure for the server and private for the clients simultaneously. Most prior efforts fall into either of the two categories. At one end of the spectrum, some work uses techniques like secure aggregation to hide the individual client’s updates and only reveal the aggregated global update to a malicious server that strives to infer the clients’ privacy from their updates. At the other end of the spectrum, some work uses Byzantine-robust FL protocols to suppress the influence of malicious clients’ updates. We present a protocol that offers bidirectional defense to simultaneously combat against the malicious centralized server and Byzantine malicious clients. Our protocol also improves the dimension dependence and achieve a near-optimal statistical rate for strongly convex cases.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31230 2024-05-20T15:06:29-07:00 Christina Alexandris calexandris@gs.uoa.gr

It is a widely-accepted fact that the processing of very large amounts of data with state-of-the-art Natural Language Processing (NLP) practices (i.e. Machine Learning –ML, language agnostic approaches) has resulted to a dramatic improvement in the speed and efficiency of systems and applications. However, these developments are accompanied with several challenges and difficulties that have been voiced within the last years. Specifically, in regard to NLP, evident improvement in the speed and efficiency of systems and applications with GenAI also entails some aspects that may be problematic, especially when particular text types, languages and/or user groups are concerned. State-of-the-art NLP approaches with automated processing of vast amounts of data in GenAI are related to observed problematic Aspects 1-7, namely: (1) Underrepresentation, (2) Standardization. These result to (3) Barriers in Text Understanding, (4) Discouragement of HCI Usage for Special Text Types and/or User Groups, (5) Barriers in Accessing Information, (6) Likelihood of Errors and False Assumptions and (7) Difficulties in Error Detection and Recovery. An additional problem are typical cases, such as less-resourced languages (A), less experienced users (B) and less agile users (C). A hybrid approach involving the re-introduction and integration of traditional concepts in state-of-the-art processing approaches, whether they are automatic or interactive, concerns the following targets: i), (ii) and (iii): Making more types of information accessible to more types of recipients and user groups (i), Making more types of services accessible and user-friendly to more types of user groups (ii), Making more types of feelings, opinions, voices and reactions visible from more types of user groups (iii) Specifically, in the above-presented cases traditional and classical theories, principles and models are re-introduced and can be integrated into state-of-the art data-driven approaches involving Machine Learning and neural networks, functioning as training data and seed data in Natural Language Processing applications where user requirements and customization are of particular interest and importance. A hybrid approach may be considered a compromise between speed and correctness / userfriendliness in (types of) NLP applications where the achievement of this balance plays a crucial role. In other words, a hybrid approach and the examples presented here target to prevent mechanisms from adopting human biases, ensuring fairness and socially responsible outcome and responsible Social Media. A hybrid approach and the examples presented here also target to customizing content to different linguistic and cultural groups, ensuring equitable information distribution. Here, we present characteristic examples with cases employing the re-introduction of four typical types of traditional concepts concerning classical theories, principles and models. These four typical classical theories, principles and models are also not considered to be flawless, however they can be transformed into practical strategies that can be integrated into evaluation modules, neural networks and training data (including knowledge graphs) and dialogue design. The proposed and discussed re-introduction of traditional concepts is not limited only to the particular models, principles and theories presented here. The first example concerns the application of a classic principle from Theoretical Linguistics. The concept employed in the second example concerns a model from the field of Linguistics and Translation. The third and the fourth examples demonstrate the interdisciplinary application of models and theoretical frameworks from the fields of Linguistics-Cognitive Science and Linguistics-Psychology respectively.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31231 2024-05-20T15:06:31-07:00 Soki Arai a2230007@gl.cc.uec.ac.jp Yuki Yamamoto y2010714@ms.cc.uec.ac.jp Yuji Nozaki na003169@uec.ac.jp Haruka Matsukura matsukura@uec.ac.jp Maki Sakamoto maki.sakamoto@uec.ac.jp

Various meetings are carried out in intellectual production activities and workers have to spend much time to create ideas. In creative meetings, it is sometime difficult for the meeting moderators and facilitators to efficiently conduct the meetings because the participants are required to come up with new ideas one after another and some participants hesitate to express unconventional ideas. Therefore, we propose to develop an AI-based meeting support system that estimates participants’ inspiration and helps to generate comfortable meeting environments for improvement of worker wellbeing. Participants’ inspiration is assumed to be estimated based on their speech and micro behaviors including smiles and nods. In this paper, a dataset we collected for the development of the proposed system is reported. The dataset consists of participants’ brain blood flows measured near-infrared spectrometers, micro behavior annotated from video recording, and inspiration the participants reported with buttons. The data for 1020 min was collected by conducting simulation meetings. In future work, we plan to train an LSTM (long short-term memory) based neural network model to realize the proposed system.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31232 2024-05-20T15:06:33-07:00 Oliver Bendel oliver.bendel@fhnw.ch

This paper examines the question of how generative AI can improve the well-being of blind or visually impaired people. It refers to a current example, the Be My Eyes app, in which the Be My AI feature was integrated in 2023, which is based on GPT-4 from OpenAI. The author’s tests are described and evaluated. There is also an ethical and social discussion. The power of the tool, which can analyze still images in an amazing way, is demonstrated. Those affected gain a new independence and a new perception of their environment. At the same time, they are dependent on the world view and morality of the provider or developer, who prescribe or deny them certain descriptions. An outlook makes it clear that the analysis of moving images will mean a further leap forward. It is fair to say that generative AI can fundamentally improve the well-being of blind and visually impaired people and will change it in various ways.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31233 2024-05-20T15:06:34-07:00 Sara Darwish Darwish_Sara@bah.com Alison Bragaw-Butler Bragaw-Butler_Alison@bah.com Paul Marcelli 68cfadf8011391af86434c58001158ec@example.org Kaylee Gassner 5da865146cd0fb3a89af268992419710@example.org

Artificial Intelligence (AI) adoption has seen substantial growth across industries. This paper explores the escalating use of AI within the United States Department of Defense (DoD) and the implications that diversity, equity, and inclusion (DEI) have on Service members and Civilians across the Department. More specifically, this paper explores the DEI considerations within AI technologies on individual, team, and Department readiness. The DoD's AI usage spans various strategic and operational capabilities, however this paper explores two critical domains: healthcare and recruitment. In healthcare, AI offers the promise of early disease detection, enhanced diagnostic capabilities, and streamlined administrative processes. However, potential biases stemming from homogenous training data threaten the accuracy and reliability of these systems, jeopardizing Service member health and eroding trust in AI-assisted medical decision-making and potentially the DoD at large. In recruitment, while AI promises efficiency in identifying ideal candidates, its deployment can perpetuate biases, especially when the training data used is not representative of all demographics. Despite efforts to design "unbiased" systems by excluding demographic data, such strategies may inadvertently overlook the unique challenges faced by marginalized communities, further entrenching existing disparities. Both case studies underscore the importance of considering DEI in the development and deployment of AI systems. As the DoD continues to integrate AI into its operations, this paper’s recommendations stress the necessity of continuous DEI assessment to ensure that AI serves as an asset rather than a liability. The authors recommend the following: 1. Data diversity & review 2. Continuous monitoring and calibration 3. Stakeholder engagement 4. Adoption of DEI requirements within Ethical AI Frameworks 5. Further research

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31234 2024-05-20T15:06:35-07:00 Stefanie Hauske haks@zhaw.ch Oliver Bendel oliver.bendel@fhnw.ch

This paper explores how generative AI (GenAI) can improve the well-being of learners within self-regulated learning (SRL) frameworks in the corporate context. In the “GenAI to Support SRL” section, it presents three custom versions of ChatGPT aimed at assisting learners. These so-called GPTs demonstrate the GenAI’s potential to actively support learners in SRL and positively influence their well-being. The “Discussion” and “Summary and Outlook” sections provide a balanced overview of the opportunities and risks associated with GenAI in the field of learning and highlight directions for future research. The results indicate that GenAI could improve the well-being of learners in SRL through providing personalized guidance, reducing feelings of stress, and increasing motivation and self-efficacy. At the same time, there are several challenges for companies and employees that need to be overcome.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31235 2024-05-20T15:06:36-07:00 Kazuhiro Ito ito.kazuhiro.ih4@is.naist.jp Junko Hayashi Junko@example.com Shoko Wakamiya Shoko@example.com Masae Manabe Manabe@example.com Yasushi Watanabe Watanabe@example.com Masataka Nakayama masataka@example.com Yukiko Uchida Yukiko@example.com Eiji Aramaki aramaki@is.naist.jp

Although well-being is helpful in measuring the state of society from various perspectives, past research has been limited to (1) questionnaire surveys, which make it difficult to target a large number of people, and (2) the major indices focus on individual factors and do not incorporate group factors. To tackle these issues, we collected daily reports from the company employees that included text, their individual subjective well-being, and team subjective well-being. By using the collected data, we constructed a well-being estimation model based on the Large Language Model and examined an indicator called ``sharedness index'', as a state of the team that influences an individual well-being, measured using both score- and text-based methods.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31236 2024-05-20T15:06:38-07:00 Takashi Kido kido.takashi@gmail.com Keiki Takadama keiki@inf.uec.ac.jp

At the AAAI Spring Symposium 2024, we explore the important challenges facing Generative Artificial Intelligence (GenAI) concerning both social structures and individual welfare. Our discussion revolves around two perspectives. Individual Impact of GenAI on Well-being: This perspective focuses on the design of AI systems with keen consideration for individual well-being. It seeks to understand how digital experiences influence emotions and the quality of life at a personal level. By examining the effects of AI technologies on individuals, we aim to tailor solutions to enhance personal welfare and fulfillment. Social Impact of GenAI on Well-being: Here, emphasis shifts to the broader societal implications of GenAI. We strive for decisions and implementations that foster fairness and benefit all members of society. This perspective acknowledges the interconnectedness of individuals within social structures and seeks to ensure that GenAI advancements positively contribute to collective well-being. In this paper, we provide an overview of the motivations driving our exploration, elucidate key terms essential for understanding the discourse, outline the primary areas of focus of our symposium, and pose research inquiries that will guide our discussions. Through this comprehensive approach, we aim to address the multifaceted challenges and opportunities presented by GenAI in promoting both social and individual well-being.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31237 2024-05-20T15:06:39-07:00 Iko Nakari iko0528@cas.lab.uec.ac.jp Keiki Takadama keiki@inf.uec.ac.jp

As a first step towards realizing an AI sleep counselor capable of generating personalized advice, this paper proposes a method for monitoring daily sleep conditions with a mattress sensor. To improve the accuracy of sleep stage estimation and to get accurate sleep structure, this paper introduced sleep domain knowledge to machine learning for improving the accuracy of sleep stage estimation. Concretely, the proposed method estimates ultradian rhythm based on the body movement density, updates prediction probabilities of each sleep stage by ML model and applies WAKE/NR3 detection based on the large/small body movement. Through the human subject experiment, the following implications have been revealed: (1) the proposed method improved the percentage of Accuracy by 65.0% from 61.5% and the QWK score by 0.196 from 0.297 by the conventional machine learning method; (2) the proposed method prevents over-NR12 estimating and is useful for understanding sleep structure by estimating REM sleep and NR3 sleep correctly. (3) the correct estimation of ultradian rhythms significantly improved the sleep stage estimation, with an Accuracy of 77.6% and a QWK score of 0.52 when all subjects' ultradian rhythms were estimated correctly.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31238 2024-05-20T15:06:41-07:00 Yuto Nakashima yuto-nakashima@g.ecc.u-tokyo.ac.jp

Generating preferred images from GANs is a challenging task due to the high-dimensional nature of latent space. In this study, we propose a novel approach that uses simple user-swipe interactions to generate preferred images from users. To effectively explore the latent space with only swipe interactions, we apply principal component analysis to the latent space of StyleGAN, creating meaningful subspaces. Additionally, we use a multi-armed bandit algorithm to decide which dimensions to explore, focusing on the user's preferences. Our experiments show that our method is more efficient in generating preferred images than the baseline.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31239 2024-05-20T15:06:42-07:00 Michelle Nie michelle.nie@sciencespo.fr

The impact of generative artificial intelligence (GenAI) on increasing misinformation is well-understood. But there remain questions on how GenAI impacts the well-being of individuals and societies at large. This paper tackles this question from a political science standpoint and considers the impact on democracy, which is linked to individual and social well-being. It examines aspects of AI systems, including GenAI systems, that threaten to undermine democracy the most, such as misinformation. This paper also clarifies the nature of these threats to democracy, makes the connection to epistemic agency and political trust, and outlines potential outcomes to society and political institutions, including accelerating the rise of populism, the enhancement of authoritarian governments, and the threat of rule by algorithms.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31240 2024-05-20T15:06:43-07:00 Mark Nuppnau hi6094@wayne.edu Khalid Kattan kkattan@umich.edu R. G. Reynolds robert.reynolds@wayne.edu

This study explores the integration of cultural algorithms (CA) with the Policy Gradients with Parameter-Based Exploration (PGPE) algorithm for the task of MNIST hand-written digit classification within the EvoJAX framework. The PGPE algorithm is enhanced by incorporating a belief space, consisting on Domain, Situational, and History knowledge sources (KS), to guide the search process and improve convergence speed. The PGPE algorithm, implemented within the EvoJAX framework, can efficiently find an optimal parameter-space policy for the MNIST task. However, increasing the complexity of the task and policy space, such as the CheXpert dataset and DenseNet, requires a more sophisticated approach to efficiently navigate the search space. We introduce CA-PGPE, a novel approach that integrates CA with PGPE to guide the search process and improve convergence speed. Future work will focus on incorporating exploratory knowledge sources and evaluate the enhanced CA-PGPE algorithm on more complex datasets and model architectures, such as CIFAR-10 and CheXpert with DenseNet.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31241 2024-05-20T15:06:44-07:00 Yukio Ohsawa ohsawa@sys.t.u-tokyo.ac.jp Tomohide Maekawa 7bd8c9a0d840eb391e22630cec607d7a@example.org Hiroki Yamaguchi bc572f48e0c9ef4f694d2ccd38c05e18@example.org Hiro Yoshida 77be042aa540954ffc0e310365012d09@example.org Kaira Sekiguchi af974470891d99d51ed1d162071bd0c6@example.org

Feature concepts and data leaves have been invented to foster thoughts for creating social and physical well-being through the use of datasets. The idea, simply put, is to at-tach selected and collected Data Leaves that are summaries of event flows to be discovered from corresponding datasets, on the target Feature Concept representing the expected scenarios of well-being individuals and well-being society. A graph of existing or expected datasets, attached in the form of Data Leaves on a Feature Concept, was generated semi-automatically. Rather than sheer auto-mated generative AI, our work addresses the process of generative artificial and natural intelligence to create the basis for collecting and connecting useful data.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31242 2024-05-20T15:06:46-07:00 Yukio Ohsawa ohsawa@sys.t.u-tokyo.ac.jp Sae Kondo skondo@arch.mie-u.ac.jp Yi Sun sun-yi650@g.ecc.u-tokyo.ac.jp Kaira Sekiguchi kaira@sys.t.u-tokyo.ac.jp

The well-being of individuals in a crowd is interpreted as a product of individuals crossing over from heterogeneous communities, via interactions with other crowds. Here, the index moving-direction entropy corresponding to the diversity of the moving directions of individuals is introduced to represent such an inter-community crossover and extended with multiscale scopes. Multiscale moving direction entropies, composed of various geographical mesh sizes to compute the index values, are used to capture the flow and interaction of information owing to human movements from/to various crowds. The generated map of high values of multiscale moving direction entropy was visualized, where the peaks coincided significantly with the preference of people to live in each region.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31243 2024-05-20T15:06:47-07:00 Yohn Jairo Parra Bautista yohn.parrabautista@famu.edu Carlos Theran carlos.theran@famu.edu Richard Aló richard.alo@famu.edu

We explore the foresighted concerns that Norbert Wiener voiced in 1960 about the potential of machines to learn and create strategies that could not be anticipated, drawing parallels to the fable "The Sorcerer's Apprentice" by Goethe. The progress in artificial intelligence (AI) has brought these worries back to the forefront, as shown by a survey AI Impacts conducted in 2022 with more than 700 machine learning researchers. This survey found a five percentage probability that advanced AI might cause "extremely adverse" outcomes, including the possibility of human extinction. Importantly, the introduction of OpenAI's ChatGPT, powered by GPT-4, has led to a surge in entrepreneurial activities, highlighting the ease of use of large language models (LLMs).AI's potential for adverse outcomes, such as military control and unregulated AI races, is explored alongside concerns about AI's role in governance, healthcare, media portrayal, and surpassing human intelligence. Given their transformative impact on content creation, the prominence of generative AI tools such as ChatGPT is noted. The societal assessment of Artificial Intelligence (AI) has grown increasingly intricate and pressing in tandem with the rapid evolution of this technology. As AI continues to advance at a swift pace, the need to comprehensively evaluate its societal implications has become more complex and urgent, necessitating a thorough examination of its potential impact on various domains such as governance, healthcare, media portrayal, and surpassing human intelligence. This assessment is crucial in addressing ethical concerns related to bias, data misuse, technical limitations, and transparency gaps, and in integrating ethical and legal principles throughout AI algorithm lifecycles to ensure alignment with societal well-being. Furthermore, the urgency of addressing the societal implications of AI is underscored by the need for healthcare workforce upskilling and ethical considerations in the era of AI-assisted medicine, emphasizing the critical importance of integrating societal well-being into the development and deployment of AI technologies. Our study entails an examination of the ethical quandaries and obstacles presented when developing methods to evaluate and predict the broader societal impacts of AI on decision-making processes involving the generating of images, videos, and textual content.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31244 2024-05-20T15:06:49-07:00 Helen Qin helen.ty.qin@gmail.com

This paper reports how generative AI can help children with specific language impairment (SLI) issues by developing an AI-assisted tool to support children with challenges in phonological development in English, especially children with English as the secondary language in the United States. Children from bilingual families often experience challenges in developing proficiency in English pronunciation and communication, which has been exacerbated by remote learning during the pandemic and led to learning loss. School-aged children with speech problems require timely intervention because children with language disorders find it difficult to communicate with others, leading to social isolation and academic difficulties. The needed intervention is often delayed due to the high cost of speech services and the shortage of Speech and Language Pathologists (SLPs). Individuals with a history of SLI have an increased risk of unemployment. An AI-assisted Phonological Development (AI-PD) tool was prototyped, aiming to alleviate these challenges by assisting caregivers in evaluating children's phonological development, assisting SLPs in lesson preparation, and mitigating the severe shortage of SLPs.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31245 2024-05-20T15:06:50-07:00 Zhonghao Shi zhonghas@usc.edu Ellen Landrum cf6ce4ade20d987dd589e21fedeca642@example.org Amy O'Connell 5b1271bf820c65784a9d69ba3f46781e@example.org Mina Kian 3ce3cb2574c9e8dd7c19f046e1b69ad3@example.org Leticia Pinto-Alva 0d95e6cbc0eb30b67c7faff585b5b9ac@example.org Kaleen Shrestha 6291b17d8735387a17e09cfc190c43cd@example.org Xiaoyuan Zhu b85d0dd73cae0a856899cfaa576a1468@example.org Maja J Matarić cdaa476899e22e7d181b875e791ed235@example.org

Socially assistive robots (SARs) have shown great success in providing personalized cognitive-affective support for user populations with special needs such as older adults, children with autism spectrum disorder (ASD), and individuals with mental health challenges. The large body of work on SAR demonstrates its potential to provide at-home support that complements clinic-based interventions delivered by mental health professionals, making these interventions more effective and accessible. However, there are still several major technical challenges that hinder SAR-mediated interactions and interventions from reaching human-level social intelligence and efficacy. With the recent advances in large language models (LLMs), there is an increased potential for novel applications within the field of SAR that can significantly expand the current capabilities of SARs. However, incorporating LLMs introduces new risks and ethical concerns that have not yet been encountered, and must be carefully be addressed to safely deploy these more advanced systems. In this work, we aim to conduct a brief survey on the use of LLMs in SAR technologies, and discuss the potentials and risks of applying LLMs to the following three major technical challenges of SAR: 1) natural language dialog; 2) multimodal understanding; 3) LLMs as robot policies.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31246 2024-05-20T15:06:51-07:00 Daiki Shintani shindaiki@cas.lab.uec.ac.jp Iko Nakari iko0528@cas.lab.uec.ac.jp Satomi Washizaki f9a0aa5a324e37ea68b6b0ca3780613f@example.org Keiki Takadama keiki@inf.uec.ac.jp

This paper proposes the method by physiological knowledge to improve the estimation performance of the NREM3 sleep based on the waist-attached accelerometer. Specifically, this paper proposes the hybrid method that combines the method based on body movement counts and the method based on biological rhythms of sleep. Through the human subject experiment, the following implications were revealed: (1) the proposed method can outperform famous machine learning models (Random Forest and LSTM) trained with automatically generated features that do not sufficiently incorporate domain knowledge; (2) when the input features are based on domain knowledge, the estimator explicitly designed by humans can outperform the machine learning method; and (3) combining the body movement counting method and the biological rhythm-based method can suppress the error of the body movement counting method and reduce false positives.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31247 2024-05-20T15:06:53-07:00 Andy Skumanich drandysku@gmail.com Han Kyul Kim hankyulk@usc.edu

A rapidly developing threat to societal well-being is from misinformation widely spread on social media. Even more concerning is ”mal-info” (malicious) which is amplified on certain social networks. Now there is an additional dimension to that threat, which is the use of Generative AI to deliberately augment the mis-info and mal-info. This paper highlights some of the ”fringe” social media channels which have a high level of mal-info as characterized by our AI/ML algorithms. We discuss various channels and focus on one in particular, ”GAB”, as representative of the potential negative impacts. We outline some of the current mal-info as an example. We capture elements, and observe the trends in time. We provide a set of AI/ML modes which can characterize the mal-info and allow for capture, tracking, and potentially for responding or for mitigation. We highlight the concern about malicious agents using GenAI for deliberate mal-info messaging specifically to disrupt societal well being. We suggest the characterizations presented as a methodology for initiating a more deliberate and quantitative approach to address these harmful aspects of social media which would adversely impact societal well being. The article highlights the potential for ”mal-info,” including disinfo, cyberbullying, and hate speech, to disrupt segments of society. The amplification of mal-info can result in serious real-world consequences such as mass shootings. Despite attempts to introduce moderation on major platforms like Facebook and to some extent on X/Twitter, there are now growing social networks such as Gab, Gettr, and Bitchute that offer completely unmoderated spaces. This paper presents an introduction to these platforms and the initial results of a semiquantitative analysis of Gab’s posts. The paper examines several characterization modes using text analysis. The paper emphasizes the developing dangerous use of generative AI algorithms by Gab and other fringe platforms, highlighting the risks to societal well being. This article aims to lay the foundation for capturing, monitoring, and mitigating these risks.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31248 2024-05-20T15:06:54-07:00 Hiroshi Sogabe h.sogabe123@gmail.com Masayuki Numao masayuki.numao@uec.ac.jp

Currently, MRI examinations and neuropsychological tests by physicians and clinical psychologists are used to screen for dementia, but they are problematic because they overwhelm medical resources and are highly invasive to patients. If automatic detection of dementia from conversations becomes feasible, it will reduce the burden on medical institutions and realize a less invasive screening method. In this paper, we constructed a machine learning model to identify dementia by extracting linguistic features and time features from the elderly corpus with a control group. Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR) were used in the model. We compared the AUC of the single topic model and the general topic model in three cases: (I) All Features, (II) Gini Impurity, and (III) PCA + Gini Impurity. The AUC of the model constructed using RF in (III) for a single topic was 0.91, showing higher AUC than in the previous study. Furthermore, topic analysis showed that topics with high similarity in utterance content are effective in identifying MCI. In the case of the general topic, the model with AUC of 0.8 showed a high identification performance for unknown topics by cross validation on a topic-by-topic basis, indicating that the general topic model developed in this study can be applied to general conversation.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31249 2024-05-20T15:06:55-07:00 Melanie Swan lfkvtz@yahoo.com Takashi Kido d658c554f3bedda77da2a1b465121341@example.org Eric Roland 5c21b5104db98b9787ba036ccce772a3@example.org Renato P. dos Santos ce6b90456f78310514f3713beb9c8a2d@example.org

Health Agents are introduced as the concept of a personalized AI health advisor overlay for continuous health monitoring (e.g. 1000x/minute) medical-grade smartwatches and wearables for “healthcare by app” instead of “sickcare by appointment.” Individuals can customize the level of detail in the information they view. Health Agents “speak” natural language to humans and formal language to the computational infrastructure, possibly outputting the mathematics of personalized homeostatic health as part of their reinforcement learning agent behavior. As an AI health interface, the agent facilitates the management of precision medicine as a service. Healthy longevity is a high-profile area characterized by the increasing acceptance of medical intervention, longevity biotech venture capital investment, and global priority as 2 billion people will be over 65 in 2050. Aging hallmarks, biomarkers, and clocks provide a quantitative measure for intervention. Some of the leading interventions include metformin, rapamycin, spermidine, NAD+/sirtuins, alpha-ketoglutarate, and taurine. AI-driven digital biology, longevity medicine, and Web3 personalized healthcare come together in the idea of Health Agents. This Web3 genAI tool for automated health management, specifically via digital-biological twins and pathway2vec approaches, demonstrates human-AI intelligence amplification and works towards healthy longevity for global well-being.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31250 2024-05-20T15:06:57-07:00 Keiki Takadama keiki@inf.uec.ac.jp

This paper explores an answer to the question of “what is a correct output by generative AI from the viewpoint of well-being?” and discusses an effectiveness of taking account of a biological rhythm for this issue. Concretely, this paper focuses on an estimation of the REM sleep stage as one of sleep stages, and compared its estimations based on random forest as one of the machine learning methods and the ultradian rhythm as one of the biological rhythms. From the human subject experiment, the following implications have been revealed: (1) the REM sleep stage is wrongly estimated in many areas by random forest; and (2) the integration of the REM sleep stage estimation based on the biological rhythm with that based on random forest improves the F-score of the estimated REM sleep stage.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31251 2024-05-20T15:06:58-07:00 Faye-Marie Vassel fvassel@stanford.edu Evan Shieh youngdatascientists@gmail.com Cassidy R. Sugimoto sugimoto@gatech.edu Thema Monroe-White tmonroewhite@berry.edu

The rapid emergence of generative Language Models (LMs) has led to growing concern about the impacts that their unexamined adoption may have on the social well-being of diverse user groups. Meanwhile, LMs are increasingly being adopted in K-20 schools and one-on-one student settings with minimal investigation of potential harms associated with their deployment. Motivated in part by real-world/everyday use cases (e.g., an AI writing assistant) this paper explores the potential psychosocial harms of stories generated by five leading LMs in response to open-ended prompting. We extend findings of stereotyping harms analyzing a total of 150K 100-word stories related to student classroom interactions. Examining patterns in LM-generated character demographics and representational harms (i.e., erasure, subordination, and stereotyping) we highlight particularly egregious vignettes, illustrating the ways LM-generated outputs may influence the experiences of users with marginalized and minoritized identities, and emphasizing the need for a critical understanding of the psychosocial impacts of generative AI tools when deployed and utilized in diverse social contexts.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31252 2024-05-20T15:06:59-07:00 Kevin Vo kvo1@ualberta.ca

In this ongoing narrative review, I summarize the existing body of literature on the role of artificial intelligence in mediating human communication, focusing on how it is currently transforming our communication patterns. Moreover, this re-view uniquely contributes by critically analyzing potential future shifts in these patterns, particularly in light of the advancing capabilities of artificial intelligence. Special emphasis is placed on the implications of emerging generative AI technologies, projecting how they might redefine the landscape of human interaction.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31253 2024-05-20T15:07:00-07:00 Kevin Vo kvo1@ualberta.ca

This study proposal combines the transformative potential of GPT-4 with an innovative approach to learning social and emotional skills, offering a novel conversational aid designed to enhance adolescents' social competence, and ultimately combat social disconnection in the digital era.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31254 2024-05-20T15:07:01-07:00 Jin Yamanaka jin.yamanaka@gmail.com Takashi Kido kido.takashi@gmail.com

After the significant performance of Large Language Models (LLMs) was revealed, their capabilities were rapidly expanded with techniques such as Retrieval Augmented Generation (RAG). Given their broad applicability and fast development, it's crucial to consider their impact on social systems. On the other hand, assessing these advanced LLMs poses challenges due to their extensive capabilities and the complex nature of social systems. In this study, we pay attention to the similarity between LLMs in social systems and humanoid robots in open environments. We enumerate the essential components required for controlling humanoids in problem solving which help us explore the core capabilities of LLMs and assess the effects of any deficiencies within these components. This approach is justified because the effectiveness of humanoid systems has been thoroughly proven and acknowledged. To identify needed components for humanoids in problem-solving tasks, we create an extensive component framework for planning and controlling humanoid robots in an open environment. Then assess the impacts and risks of LLMs for each component, referencing the latest benchmarks to evaluate their current strengths and weaknesses. Following the assessment guided by our framework, we identified certain capabilities that LLMs lack and concerns in social systems.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31255 2024-05-20T15:07:02-07:00 Mingzhe Yang mingzhe-yang@g.ecc.u-tokyo.ac.jp

Human judgment is often subject to bias, leading to unfair decisions. This is particularly problematic when assessments have significant consequences, underscoring the importance of guiding humans towards fairness. Although recent advancements in AI have facilitated decision support, it is not always feasible to employ AI assistance in real-world scenarios. Therefore, this study focuses on developing and evaluating a method to guide humans in making fair judgments. Our experimental results confirmed that our approach effectively promotes fairness in human decision-making.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31256 2024-05-20T15:07:03-07:00 Sharon Chee Yin Ho sharoncheeyin.ho@concordia.ca Arisa Ema 6463e7762ab5df486595d59aeb6cf209@example.org Tanja Tajmel 8253a00a04d400c23006491432cda9e4@example.org

The growing interest in Japan to implement text-to-image (T2I) generative artificial intelligence (GenAI) technologies in creative workflows has raised concern over what ethical and social implications these technologies will have on creative professionals. Our pilot study is the first to discuss what social and ethical oversights may emerge regarding such issues from prospective Japanese researchers – computer science (CS) graduate students studying in Japan. Given that these students are the primary demographic hired to work at research and development (R&D) labs at the forefront of such innovations in Japan, any social and ethical oversight on such issues may unequip them as future knowledge experts who will play a pivotal role in helping shape Japan’s policies regarding image generating AI technologies.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31257 2024-05-20T15:07:04-07:00 Takahiro Yonekawa yonekawa@bsgnl.com Hiroko Yamano yamano@ifi.u-tokyo.ac.jp Ichiro Sakata 3f26f6f1827e96cf3f7edd1c19b57286@example.org

Interactive generative AI can be used in software programming to generate sufficient quality of code. Software developers can utilize the output code of generative AI as well as website resources from search engine results. In this research, we present a framework for defining states of programming activity and for capturing the actions of developers in a time series. We also describe a scheme for analyzing the thought process of software developers by using a graph structure to describe state transitions. By applying these means, we showed that it is feasible to analyze the effects of changes in the development environment on programming activities.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31258 2024-05-20T15:07:08-07:00 Sambit Bhattacharya sbhattac@uncfsu.edu Bogdan Czejdo bczejdo@uncfsu.edu Rebecca A. Zulli rebecca.anne.zulli@gmail.com Adrienne A. Smith adrienne.ann.smith@gmail.com

While students are often passionate about their chosen fields, they often have limited awareness of the profound impact of AI technologies on their professions. In order to advance efforts in building subject-relevant AI literacy among undergraduate students studying Computer Science and non-Computer Science (Criminal Justice and Forensic Science) it is imperative to engage in rigorous efforts to develop and study curricular infusion of Artificial Intelligence topics. Using a Design-Based Research model, the project team and the external evaluators studied the first iteration of the module development and implementation. Using data collected through surveys, focus groups, critical review, and reflection exercises the external evaluation team produced findings that informed the project team in revising and improving their materials and approach for the second iteration. These efforts can help educators and the AI module developers tailor their AI curriculum to address these specific areas, ensuring that students develop a more accurate understanding of applications of AI in their future career field.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31259 2024-05-20T15:07:09-07:00 Yu Chen yu.chen@sjsu.edu Gabriel Granco ggranco@cpp.edu Yunfei Hou yunfei.hou@csusb.edu Heather Macias heather.macias@csulb.edu Frank A. Gomez fgomez@calstate.edu

This project aims to broaden AI education by developing and studying the efficacy of innovative learning practices and resources for AI education for social good. We have developed three AI learning modules for students to: 1) identify social issues that align with the SDGs in their community (e.g., poverty, hunger, quality education); 2) learn AI through hands-on labs and business applications; and 3) create AI-powered solutions in teams to address social is-sues they have identified. Student teams are expected to situate AI learning in their communities and contribute to their communities. Students then use the modules to en-gage in an interdisciplinary approach, facilitating AI learn-ing for social good in informational sciences and technology, geography, and computer science at three CSU HSIs (San Jose State University, Cal Poly Pomona and CSU San Bernardino). Finally, we aim to evaluate the efficacy and impact of the proposed AI teaching methods and activities in terms of learning outcomes, student experience, student engagement, and equity.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31260 2024-05-20T15:07:10-07:00 Ryan Evans revans@wanaqueps.org Neelu Sinha sinha@fdu.edu

This position paper highlights the critical need to enhance diversity in artificial intelligence (AI) education, focusing on K-8 students. As AI increasingly shapes our societal landscape, ensuring equitable access and participation in AI-related fields is essential. However, the current AI education landscape lacks inclusivity, resulting in underrepresentation and limited opportunities for marginalized groups such as racial and ethnic minorities, women, individuals with disabilities, and those from economically disadvantaged backgrounds. The paper advocates for a comprehensive approach to address diversity gaps in AI education. This involves revising curricula to include diverse perspectives, integrating AI knowledge into core subject areas, and utilizing machine learning (ML) to enhance learning across disciplines. Educators can create inclusive learning environments by incorporating culturally relevant examples and interactive activities showcasing AI's positive impact on diverse communities. Furthermore, promoting diversity in AI education requires investment in teacher training and resources. Educators need support to implement inclusive teaching methods, understand cultural nuances, and address implicit biases. Bridging the digital gap is also crucial, as access to technology and hands-on AI experience ensures equal opportunities for all students regardless of socioeconomic back-ground. By embracing diversity and inclusivity in AI education at the K-8 level, we can cultivate a future generation of AI professionals and informed citizens who leverage technology to address diverse community needs.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31261 2024-05-20T15:07:11-07:00 Derek Jacoby derekja@gmail.com Saiph Savage s.savage@northeastern.edu Yvonne Coady ycoady@gmail.com

Increasing diversity in educational settings is challenging in part due to the lack of access to resources for non-traditional learners in remote communities. Post-pandemic platforms designed specifically for remote and hybrid learning---supporting team-based collaboration online---are positioned to bridge this gap. Our work combines the use of these new platforms with co-creation and collaboration tools for AI assisted remote Work-Integrated-Learning (WIL) opportunities, including efforts in community and with the public library system. This paper outlines some of our experiences to date, and proposes methods to further integrate AI education into community-driven applications for remote WIL.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31262 2024-05-20T15:07:13-07:00 Devang Jayachandran dxj5305@psu.edu Pranit Maldikar ppm5404@psu.edu Tyler S. Love tslove@umes.edu Jeremy J. Blum jjb24@psu.edu

Generative Artificial Intelligence (AI) was incorporated into a competitive programming event that targeted undergraduate students, including those with little programming experience. The competition incorporated a range of challenge design approaches that promoted meaningful interaction with generative AI system, even while keeping the challenge difficulty level to an appropriate level. An analysis of survey responses and competition data showed that this format lowered barriers to participation, successfully engaged students throughout the competition, and increased the likelihood that they would participate in a similar event. In an extension of this work, a professional development workshop for high school teachers is being developed, along with a contest for high school students. Participant surveys and logs of interaction with the contest and generative AI systems will be analyzed to measure the effect of generative AI on student self-efficacy and suggest ways to integrate generative AI instruction into computer science curriculum.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31263 2024-05-20T15:07:14-07:00 Mary Lou Maher marylou.maher@gmail.com Sri Yash Tadimalla stadimal@uncc.edu

AI is rapidly emerging as a tool that can be used by everyone, increasing its impact on our lives, society, and the economy. There is a need to develop educational programs and curricula that can increase capacity and diversity in AI as well as awareness of the implications of using AI-driven technologies. This paper reports on a workshop whose goals include developing guidelines for ensuring that we expand the diversity of people engaged in AI while expanding the capacity for AI curricula with a scope of content that will reflect the competencies and needs of the workforce. The scope for AI education included K-Gray and considered AI knowledge and competencies as well as AI literacy (including responsible use and ethical issues). Participants discussed recommendations for metrics measuring capacity and diversity as well as strategies for increasing capacity and diversity at different level of education: K-12, undergraduate and graduate Computer Science (CS) majors and non-CS majors, the workforce, and the public.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31264 2024-05-20T15:07:15-07:00 Swati Mehrotra swati.mehrotra@gmail.com Neelu Sinha sinha@fdu.edu

Artificial Intelligence (AI) has become pervasive in modern lives, with AI generative tools driving further transformation. However, a notable issue persists: the underrepresentation of females and individuals from ethnic and racial minorities in the tech industry. Despite generally positive attitudes toward technology among young students, this enthusiasm often does not extend to aspirations for careers in the field. To address this disparity, many schools in the United States are now offering computer science and AI courses at the high school level. Nevertheless, students from underrepresented groups often feel disconnected from these subjects, leading to low enrollment rates. Research underscores that students' career aspirations are solidified between the ages of 10-14 yrs, highlighting the importance of engaging them with computer science and computing skills during this formative period. Leveraging the Bourdieusian concept of social capital, this paper proposes educational interventions tailored for elementary schools. By nurturing students' technical social capital, these interventions aim to foster an inclusive ecosystem from an early age, when aspirations are taking shape. Ultimately, the goal is to enhance the accessibility of computer science education and related skills, empowering young students from underrepresented groups to pursue higher studies and careers in computer science and AI fields.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31265 2024-05-20T15:07:16-07:00 Brian Plancher bplancher@barnard.edu Sebastian Buttrich sebastian@itu.dk Jeremy Ellis keyfreemusic@gmail.com Neena Goveas neena@goa.bits-pilani.ac.in Laila Kazimierski lailakazimierski@gmail.com Jesus Lopez Sotelo jalopez@uao.edu.co Milan Lukic milan_lukic@uns.ac.rs Diego Mendez diego-mendez@javeriana.edu.co Rosdiadee Nordin rosdiadee@gmail.com Andres Oliva Trevisan olivaandres93@gmail.com Massimo Pavan massimo.pavan@polimi.it Manuel Roveri manuel.roveri@polimi.it Marcus Rüb marcus.rueb@hahn-schickard.de Jackline Tum jacklinetum17@gmail.com Marian Verhelst marian.verhelst@kuleuven.be Salah Abdeljabar salah.abdeljabar@kaust.edu.sa Segun Adebayo segun.adebayo@bowen.edu.ng Thomas Amberg thomas.amberg@fhnw.ch Halleluyah Aworinde aworinde.halleluyah@bowen.edu.ng José Bagur jabagur@uvg.edu.gt Gregg Barrett gregg.barrett@cirrusai.net Nabil Benamar n.benamar@aui.ma Bharat Chaudhari bharat.chaudhari@mitwpu.edu.in Ronald Criollo rrcrioll@espol.edu.ec David Cuartielles david.cuartielles@mau.se Jose Alberto Ferreira Filho jose.alb@unifei.edu.br Solomon Gizaw solomong@aau.edu.et Evgeni Gousev egousev@qti.qualcomm.com Alessandro Grande alessandro@edgeimpulse.com Shawn Hymel shawn@edgeimpulse.com Peter Ing peter.ing@outlook.com Prashant Manandhar prashant@infocomtech4dev.org Pietro Manzoni pmanzoni@disca.upv.es Boris Murmann bmurmann@hawaii.edu Eric Pan ep@seeed.cc Rytis Paskauskas rytis.paskauskas@ictp.it Ermanno Pietrosemoli ermanno@ictp.it Tales Pimenta tales@unifei.edu.br Marcelo Rovai rovai@unifei.edu.br Marco Zennaro mzennaro@ictp.it Vijay Janapa Reddi vj@eecs.harvard.edu

Embedded machine learning (ML) on low-power devices, also known as "TinyML," enables intelligent applications on accessible hardware and fosters collaboration across disciplines to solve real-world problems. Its interdisciplinary and practical nature makes embedded ML education appealing, but barriers remain that limit its accessibility, especially in developing countries. Challenges include limited open-source software, courseware, models, and datasets that can be used with globally accessible heterogeneous hardware. Our vision is that with concerted effort and partnerships between industry and academia, we can overcome such challenges and enable embedded ML education to empower developers and researchers worldwide to build locally relevant AI solutions on low-cost hardware, increasing diversity and sustainability in the field. Towards this aim, we document efforts made by the TinyML4D community to scale embedded ML education globally through open-source curricula and introductory workshops co-created by international educators. We conclude with calls to action to further develop modular and inclusive resources and transform embedded ML into a truly global gateway to embedded AI skills development.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31266 2024-05-20T15:07:19-07:00 Christelle Scharff cscharff@pace.edu James Brusseau jbrusseau@pace.edu Krishna Mohan Bathula kbathula@pace.edu Kaleemunnisa Fnu klnu@pace.edu Samyak Rakesh Meshram smeshram@pace.edu Om Gaikhe ogaikhe@pace.edu

This paper addresses the ethics of inclusion in artificial in-telligence in the context of African fashion. Despite the proliferation of fashion-related AI applications and da-tasets global diversity remains limited, and African fash-ion is significantly underrepresented. This paper docu-ments two use-cases that enhance AI's inclusivity by in-corporating sub-Saharan fashion elements. The first case details the creation of a Senegalese fashion dataset and a model for classifying traditional apparel using transfer learning. The second case investigates African wax textile patterns generated through generative adversarial net-works (GANs), specifically StyleGAN architectures, and machine learning diffusion models. Alongside the practi-cal, technological advances, theoretical ethical progress is made in two directions. First, the cases are used to elabo-rate and define the ethics of inclusion, while also contrib-uting to current debates about how inclusion differs from ethical fairness. Second, the cases engage with the ethical debate on whether AI innovation should be slowed to prevent ethical imbalances or accelerated to solve them.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31267 2024-05-20T15:07:21-07:00 Neelu Sinha sinha@fdu.edu Rama Madhavarao ramam@fdu.edu Robert Freeman robertf@student.fdu.edu Irene Oujo oujo@fdu.edu Janet Boyd boydj@fdu.edu

Degree completion rates for Hispanic students lag far be-hind their white non-Hispanic peers. To close this gap and accelerate degree completion for Hispanic students at Hispanic-Serving Institutions (HSIs), we offer a pedagogical framework to incorporate AI Literacy into existing programs and encourage faculty-mentored undergraduate research initiatives to solve real-world problems using AI. Using a holistic perspective that includes experience, perception, cognition, and behavior, we describe the ideal process of learning based on a four-step cycle of experience, reflecting, thinking, and acting. Additionally, we emphasize the role of social interaction and community in developing mental abilities and understand how cognitive development is influenced by cultural and social factors. Tailoring the content to be culturally relevant, accessible, and engaging to our Hispanic students, and employing projects-based learning, we offer hands-on activities based on social justice, inclusion, and equity to incorporate AI Literacy. Furthermore, combining the pedagogical framework along with faculty-mentored undergraduate research (the significance of which has been shown to have numerous benefits) will enable our Hispanic students develop competencies to critically evaluate AI technologies, communicate and collaborate effectively with AI, and use AI as a tool anywhere; preparing them for the future and encouraging them to use AI ethically.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31268 2024-05-20T15:07:22-07:00 Sri Yash Tadimalla stadimal@uncc.edu Mary Lou Maher mmaher9@uncc.edu

The field of Artificial Intelligence (AI) is rapidly advancing, with significant potential to transform society. However, it faces a notable challenge: lack of diversity, a longstanding issue in STEM fields. In this context, this position paper examines the intersection of AI and identity as a pathway to understanding biases, inequalities, and ethical considerations in AI development and deployment. We present a multifaceted definition of AI identity, which encompasses its creators, applications, and their broader impacts. Understanding AI's identity involves analyzing the diverse individuals involved in AI's development, the technologies produced, and the social, ethical, and psychological implications. After exploring the AI identity ecosystem and its societal dynamics, We propose a framework that highlights the need for diversity in AI across three dimensions: Creators, Creations, and Consequences through the lens of identity. This paper presents a research framework for examining the implications and changes needed to foster a more inclusive and responsible AI ecosystem through the lens of identity.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31269 2024-05-20T15:07:23-07:00 Kathleen Timmerman kathleen.timmerman@uky.edu Judy Goldsmith goldsmit@uky.edu Brent Harrison brent.harrison@uky.edu Zongming Fei zongming.fei@uky.edu

For decades, the route to familiarity in AI was through technical studies such as computer science. Yet AI has infiltrated many areas of our society. Many fields are rightfully now demanding at least a passing familiarity with machine learning: understanding the standard architectures, knowledge on how to use them, and addressing common concerns. A few such fields look at the standard ethical issues such as fairness, accountability, and transparency. Very few fields situate AI technologies in sociotechnical system analysis, nor give a rigorous foundation in ethical analysis applied to the design, development, and use of the technologies. We have proposed an undergraduate certificate in AI that gives equal weight to social and ethical issues and to technical matters of AI system design and use, aimed at students outside of the traditional AI-related disciplines. By including social and ethical issues in our AI certificate requirements, we expect to attract a broader population of students. By creating an accessible AI certification, we create an opportunity for individuals from diverse experiences to contribute to the discussion of what AI is, what its impact is, and where it should go in the future.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31270 2024-05-20T15:07:27-07:00 Brendan Conway-Smith brendan.conwaysmith@carleton.ca Robert L. West robert.west@carleton.ca

Large Language Models (LLMs) lack robust metacognitive learning abilities and depend on human-provided algorithms and prompts for learning and output generation. Metacognition involves processes that monitor and enhance cognition. Learning how to learn - metacognitive learning - is crucial for adapting and optimizing learning strategies over time. Although LLMs possess limited metacognitive abilities, they cannot autonomously refine or optimize these strategies. Humans possess innate mechanisms for metacognitive learning that enable at least two unique abilities: discerning which metacognitive strategies are best and automatizing learning strategies. These processes have been effectively modeled in the ACT-R cognitive architecture, providing insights on a path toward greater learning autonomy in AI. Incorporating human-like metacognitive learning abilities into AI could potentially lead to the development of more autonomous and versatile learning mechanisms, as well as improved problem-solving capabilities and performance across diverse tasks.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31271 2024-05-20T15:07:28-07:00 Yinuo Du yinuod@andrew.cmu.edu Baptiste Prebot bprebot@andrew.cmu.edu Cleotilde Gonzalez coty@cmu.edu

During the past decade, researchers of behavioral cyber security have created cognitive agents that are able to learn and make decisions in dynamic environments in ways that assimilate human decision processes. However, many of these efforts have been limited to simple detection tasks and represent basic cognitive functions rather than a whole set of cognitive capabilities required in dynamic cyber defense scenarios. Our current work aims at advancing the development of cognitive agents that learn and make defense-dynamic decisions during cyber attacks by intelligent attack agents. We also aim to evaluate the capability of these cognitive models in ``Turing-like'' experiments, comparing the decisions and performance of these agents against human cyber defenders. In this paper, we present an initial demonstration of a cognitive model of the defender that relies on a cognitive theory of dynamic decision-making, Instance-Based Learning Theory (IBLT); we also demonstrate the execution of the same defense task by human defenders. We rely on OpenAI Gym and CybORG and adapt an existing CAGE scenario to generate a simulation experiment using an IBL defender. We also offer a new Interactive Defense Game (IDG), where \textit{human} defenders can perform the same CAGE scenario simulated with the IBL model. Our results suggest that the IBL model makes decisions against two intelligent attack agents that are similar to those observed in a subsequent human experiment. We conclude with a description of the cognitive foundations required to build autonomous intelligent cyber defense agents that can collaborate with humans in autonomous cyber defense teams.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31272 2024-05-20T15:07:30-07:00 Kenneth D. Forbus forbus@northwestern.edu

There is ample psychological evidence that analogy is ubiquitous in human learning, suggesting that computational models of analogy can play important roles in AI systems that learn in human-like ways. This talk will provide evidence for this, focusing mostly on recent advances in hierarchical analogical learning and working-memory analogical generalizations.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31273 2024-05-20T15:07:31-07:00 Matt Jones mcj@colorado.edu Tyler R. Scott tylersco@google.com Michael C. Mozer mcmozer@google.com

Natural environments have correlations at a wide range of timescales. Human cognition is tuned to this temporal structure, as seen by power laws of learning and memory, and by spacing effects whereby the intervals between repeated training data affect how long knowledge is retained. Machine learning is instead dominated by batch iid training or else relatively simple nonstationarity assumptions such as random walks or discrete task sequences. The main contributions of our work are: (1) We develop a Bayesian model formalizing the brain's inductive bias for temporal structure and show our model accounts for key features of human learning and memory. (2) We translate the model into a new gradient-based optimization technique for neural networks that endows them with human-like temporal inductive bias and improves their performance in realistic nonstationary tasks. Our technical approach is founded on Bayesian inference over 1/f noise, a statistical signature of many natural environments with long-range, power law correlations. We derive a new closed-form solution to this problem by treating the state of the environment as a sum of processes on different timescales and applying an extended Kalman filter to learn all timescales jointly. We then derive a variational approximation of this model for training neural networks, which can be used as a drop-in replacement for standard optimizers in arbitrary architectures. Our optimizer decomposes each weight in the network as a sum of subweights with different learning and decay rates and tracks their joint uncertainty. Thus knowledge becomes distributed across timescales, enabling rapid adaptation to task changes while retaining long-term knowledge and avoiding catastrophic interference. Simulations show improved performance in environments with realistic multiscale nonstationarity. Finally, we present simulations showing our model gives essentially parameter-free fits of learning, forgetting, and spacing effects in human data. We then explore the analogue of human spacing effects in a deep net trained in a structured environment where tasks recur at different rates and compare the model's behavioral properties to those of people.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31274 2024-05-20T15:07:32-07:00 Steven Jones scijones@umich.edu Peter Lindes peter.lindes@cic.iqmri.org

Human-like learning includes an ability to learn concepts from a stream of embodiment sensor data. Echoing previous thoughts such as those from Barsalou that cognition and perception share a common representation system, we suggest an addendum to the common model of cognition. This addendum poses a simultaneous semantic memory and perception learning that bypasses working memory, and that uses parallel processing to learn concepts apart from deliberate reasoning. The goal is to provide a general outline for how to extend a class of cognitive architectures to implement a more human-like interface between cognition and embodiment of an agent, where a critical aspect of that interface is that it is dynamic because of learning.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31275 2024-05-20T15:07:34-07:00 Peter Lindes peter.lindes@cic.iqmri.org Steven Jones steven.jones@cic.iqmri.org

Humans acquire language and related concepts in a trajectory over a lifetime. Concepts for simple interaction with the world are learned before language. Later, words are learned to name these concepts along with structures needed to represent larger meanings. Eventually, language advances to where it can drive the learning of new concepts. Throughout this trajectory a language processing capability uses architectural mechanisms to process language using the knowledge already acquired. We assume that this growing body of knowledge is made up of small units of form-meaning mapping that can be composed in many ways, suggesting that these units are learned incrementally from experience. In prior work we have built a system to comprehend human language within an autonomous robot using knowledge in such units developed by hand. Here we propose a research program to develop the ability of an artificial agent to acquire this knowledge incrementally and autonomously from its experience in a similar trajectory. We then propose a strategy for evaluating this human-like learning system using a large benchmark created as a tool for training deep learning systems. We expect that our human-like learning system will produce better task performance from training on only a small subset of this benchmark.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31276 2024-05-20T15:07:35-07:00 Maya Malaviya maya.anu.malaviya@gmail.com Ilia Sucholutsky is2961@princeton.edu Thomas L. Griffiths tomg@princeton.edu

What is the mechanism behind people's remarkable ability to learn from very little data, and what are its limits? Preliminary evidence suggests people can infer categories from extremely sparse data, even when they have fewer labeled examples than categories. However, the mechanisms behind this learning process are unclear. In our experiment, people learned 8 categories defined over a 2D manifold from just 4 labeled examples. Our results suggest that people are forming rich representations of the underlying categories despite this limited information. These results push the limits of how little information people need to build strong and systematic category representations.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31277 2024-05-20T15:07:36-07:00 Maya Malaviya maya.anu.malaviya@gmail.com Mark K. Ho mho4@stevens.edu

Humans are remarkably adaptive instructors who adjust advice based on their estimations about a learner’s prior knowledge and current goals. Many topics that people teach, like goal-directed behaviors, causal systems, categorization, and time-series patterns, have an underlying commonality: they map inputs to outputs through an unknown function. This project builds upon a Gaussian process (GP) regression model that describes learner behavior as they search the hypothesis space of possible underlying functions to find the one that best fits their current data. We extend this work by implementing a teacher model that reasons about a learner’s GP regression in order to provide specific information that will help them form an accurate estimation of the function.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31278 2024-05-20T15:07:37-07:00 Sheikh Mannan sheikh.mannan@colostate.edu Vivekanand Pandey Vimal somde@brandeis.edu Paul DiZio dizio@brandeis.edu Nikhil Krishnaswamy nkrishn87@gmail.com

In this paper, we explore how humans and AIs trained to perform a virtual inverted pendulum (VIP) balancing task converge and differ in their learning and performance strategies. We create a visual analogue of disoriented IP balancing, as may be experienced by pilots suffering from spatial disorientation, and train AI models on data from human subjects performing a real-world disoriented balancing task. We then place the trained AI models in a dyadic human-in-the-loop (HITL) training setting. Episodes in which human subjects disagreed with AI actions were logged and used to fine-tune the AI model. Human subjects then performed the task while being given guidance from pretrained and dyadically fine-tuned versions of an AI model. We examine the effects of HITL training on AI performance, AI guidance on human performance, and the behavior patterns of human subjects and AI models during task performance. We find that in many cases, HITL training improves AI performance, AI guidance improves human performance, and after dyadic training the two converge on similar behavior patterns.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31279 2024-05-20T15:07:39-07:00 Shiwali Mohan shiwali.mohan@gmail.com John E. Laird laird@umich.edu

Autonomous intelligent agents, including humans, operate in a complex, dynamic environment that necessitates continuous learning. We revisit our thesis that proposes that learning in human-like agents can be categorized into two levels: Level 1 (L1) involving innate and automatic learning mechanisms, while Level 2 (L2) comprises deliberate strategies controlled by the agent. Our thesis draws from our experiences in building artificial agents with complex learning behaviors, such as interactive task learning and open-world learning.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31280 2024-05-20T15:07:40-07:00 Kanad Pardeshi kpardesh@andrew.cmu.edu Itai Shapira itaishapira@g.harvard.edu Ariel Procaccia arielpro@seas.harvard.edu Aarti Singh aarti@cs.cmu.edu

Decision-making and reaching consensus are an integral part of everyday life, and studying how individuals reach these decisions is an important problem in psychology, economics, and social choice theory. Our work develops methods and theory for learning the nature of decisions reached upon by individual decision makers or groups of individuals using data. We consider two tasks, where we have access to data on: 1) Cardinal utilities for d individuals with cardinal consensus values that the group or decision maker arrives at, 2) Cardinal utilities for d individuals for pairs of actions, with ordinal information about the consensus, i.e., which action is better according to the consensus. Under some axioms of social choice theory, the set of possible decision functions reduces to the set of weighted power means, M(u, w, p) = (∑ᵢ₌₁ᵈ wᵢ uᵢᵖ)¹ᐟᵖ, where uᵢ indicate the d utilities, w ∈ ∆_{d - 1} denotes the weights assigned to the d individuals, and p ∈ ℝ (Cousins 2023). For instance, p = 1 corresponds to a weighted utilitiarian function, and p = -∞ is the egalitarian welfare function. Our goal is to learn w ∈ ∆_{d - 1} and p ∈ ℝ for the two tasks given data. The first task is analogous to regression, and we show that owing to the monotonicity in w and p (Qi 2000}, learning these parameters given cardinal utilities and social welfare values is a PAC-learnable task. For the second task, we wish to learn w, p such that, given pairs of actions u, v ∈ ℝ₊ᵈ, the preference is given as C((u, v), w, p) = sign(ln(M(u, w, p)) - ln(M(v, w, p))). This is analogous to classification; however, convexity of the loss function in w and p is not guaranteed. We analyze two related cases - one in which the weights w are known, and another in which the weights are unknown. We prove that both cases are PAC-learnable given positive u, v by giving an O(log d) bound on the VC dimension for the known weights case, and an O(d log d) bound for the unknown weights case. We also establish PAC-learnability for noisy data under IID (Natarajan 2013) and logistic noise models for this task. Finally, we demonstrate how simple algorithms can be useful to learn w and p up to moderately high d through experiments on simulated data.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31281 2024-05-20T15:07:41-07:00 Viraj Parimi vparimi@mit.edu Sungkweon Hong sungkweon5050@gmail.com Brian Williams williams@csail.mit.edu

Deep Reinforcement Learning (DRL) has garnered substantial acclaim for its versatility and widespread applications across diverse domains. Aligned with human-like learning, DRL is grounded in the fundamental principle of learning from interaction, wherein agents dynamically adjust behavior based on environmental feedback in the form of rewards. This iterative trial-and-error process, mirroring human learning, underscores the importance of observation, experimentation, and feedback in shaping understanding and behavior. DRL agents, trained to navigate complex surroundings, refine their knowledge through hierarchical and abstract representations, empowered by deep neural networks. These representations enable efficient handling of long-horizon tasks and flexible adaptation to novel situations, akin to the human ability to construct mental models for comprehending complex concepts and predicting outcomes. Hence, abstract representation building emerges as a critical aspect in the learning processes of both artificial agents and human learners, particularly in long-horizon tasks. Furthermore, human decision-making, deeply rooted in evolutionary history, exhibits a remarkable capacity to balance the tradeoff between risk and cost across various domains. This cognitive process involves assessing potential negative consequences, evaluating factors such as the likelihood of adverse outcomes, severity of potential harm, and overall uncertainty. Humans intuitively gauge inherent risks and adeptly weigh associated costs, extending beyond monetary expenses to include time, effort, and opportunity costs. The nuanced ability of humans to consider the tradeoff between risk and cost highlights the complexity and adaptability of human decision-making, a skill lacking in typical DRL agents. Principles like these derived from human-like learning present an avenue for inspiring advancements in DRL, fostering the development of more adaptive and intelligent artificial agents. Motivated by these observations and focusing on practical challenges in robotics, our efforts target risk-aware stochastic sequential decision-making problem which is crucial for tasks with extended time frames and varied strategies. A novel integration of model-based conditional planning with DRL is proposed, inspired by hierarchical techniques. This approach breaks down complex tasks into manageable subtasks(motion primitives), ensuring safety constraints and informed decision-making. Unlike existing methods, our approach addresses motion primitive improvement iteratively, employing diverse prioritization functions to guide the search process effectively. This risk-bounded planning algorithm seamlessly integrates conditional planning and motion primitive learning, prioritizing computational efforts for enhanced efficiency within specified time limits.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31282 2024-05-20T15:07:43-07:00 Edward M. Pogossian epogossi@aua.am

Our model of cognizing roots in developmental psychology by Jean Piaget, follows researchers in modeling cognizing by solvers of combinatorial games, enriches object–oriented representatives of realities by input classifiers and relationships in English, while tends to be consistent with questioning the origination of cognizing in nature. Let us introduce the basics of the model, provide arguments for its adequacy, followed by those supporting the origination of cognizing.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31283 2024-05-20T15:07:44-07:00 Ming Qian qianmi@gmail.com Chuiqing Kong chuiqingkong17@gmail.com

Our study utilizes concept description instructions and few-shot learning examples to examine the effectiveness of a large language model (GPT-4) in generating Chinese-to-English translations that embody related translation concepts. We discovered that human language experts possess superior abductive reasoning skills compared to GPT-4. Therefore, it is crucial for humans to employ abductive reasoning to craft more detailed instructions and infuse additional logic into exemplary prompts, a step essential for guiding a large language model effectively, in contrast to the more intuitive understanding a human expert might have. This approach would make the prompt engineering process more complicated and less human-like. Emphasizing domain-specific abductive reasoning stands out as a crucial aspect of human-like learning that AI/ML systems based on large language models should aim to replicate.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31284 2024-05-20T15:07:45-07:00 Irina Rabkina irina.rabkina@barnstormresearch.com

Neurotypical adult humans are impeccably good social reasoners. Despite the occasional faux pas, we know how to interact in most social settings and how to consider others' points of view. Young children, on the other hand, do not. Social reasoning, like many of our most important skills, is learned. Much like human children, AI agents are not good social reasoners. While some algorithms can perform some aspects of social reasoning, we are a ways off from AI that can interact naturally and appropriately in the broad range of settings that people can. In this talk, I will argue that learning social reasoning via the same processes used by people will help AI agents reason--and interact--more like people do. Specifically, I will argue that children learn social reasoning via analogy, and that AI agents should, too. I will present evidence from cognitive modeling experiments demonstrating the former and AI experiments demonstrating the latter. I will also propose future directions for social reasoning research that both demonstrate the need for robust, human-like social reasoning in AI and test the utility of common approaches.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31285 2024-05-20T15:07:47-07:00 Christopher B. Rauch cr625@drexel.edu Ursula Addison uaddison@parallaxresearch.org Michael Floyd michael.floyd@knexusresearch.com Prateek Goel pg427@drexel.edu Justin Karneeb justin.karneeb@knexusresearch.com Ray Kulhanek ray.kulhanek@parallaxresearch.org Othalia Larue othalia.larue@parallaxresearch.org David Ménager david.menager@parallaxresearch.org Mallika Mainali mm5579@drexel.edu Matthew Molineaux matthew.molineaux@parallaxresearch.org Adam Pease adam.pease@parallaxresearch.org Anik Sen as5867@drexel.edu Jt Turner jt.turner@knexusresearch.com Rosina Weber rw37@drexel.edu

We present an approach to algorithmic decision-making that emulates key facets of human decision-making, particularly in scenarios marked by expert disagreement and ambiguity. Our system employs a case-based reasoning framework, integrating learned experiences, contextual factors, probabilistic reasoning, domain-specific knowledge, and the personal traits of decision-makers. A primary aim of the system is to articulate algorithmic decision-making as a human-comprehensible reasoning process, complete with justifications for selected actions.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31286 2024-05-20T15:07:48-07:00 Sina Rismanchian srismanc@uci.edu Shayan Doroudi doroudis@uci.edu Yasaman Razeghi yrazeghi@uci.edu

While object recognition is one of the prevalent affordances of humans' perceptual systems, even human infants can prioritize a place system over the object recognition system, that is used when navigating. This ability, combined with active learning strategies can make humans fast learners of Turtle Geometry, a notion introduced about four decades ago. We contrast humans' performances and learning strategies with large visual language models (LVLMs) and as we show, LVLMs fall short of humans in solving Turtle Geometry tasks. We outline different characteristics of human-like learning in the domain of Turtle Geometry that are fundamentally unparalleled in state-of-the-art deep neural networks and can inform future research directions in the field of artificial intelligence.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31287 2024-05-20T15:07:49-07:00 Jesse Roberts jesse.tn.roberts@gmail.com

Human-like learning refers to the learning done in the lifetime of the individual. However, the architecture of the human brain has been developed over millennia and represents a long process of evolutionary learning which could be viewed as a form of pre-training. Large language models (LLMs), after pre-training on large amounts of data, exhibit a form of learning referred to as in-context learning (ICL). Consistent with human-like learning, LLMs are able to use ICL to perform novel tasks with few examples and to interpret the examples through the lens of their prior experience. I examine the constraints which typify human-like learning and propose that LLMs may learn to exhibit human-like learning simply by training on human generated text.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31288 2024-05-20T15:07:50-07:00 Jonathan C.H. Tong d03323003@ntu.edu.tw Yung-Fong Hsu yfhsu@ntu.edu.tw Churn-Jung Liau liaucj@iis.sinica.edu.tw

This short paper is the status report of a project in progress. We aim to model human-like agents' decision-making behaviors under risks with neural-symbolic approach. Our model integrates the learning, reasoning, and emotional aspects of an agent and takes the dual process thinking into consideration when the agent is making a decision. The model construction is based on real behavioral and brain imaging data collected in a lottery gambling experiment. We present the model architecture including its main modules and the interactions between them.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31289 2024-05-20T15:07:52-07:00 Daniel Weitekamp weitekamp@cmu.edu

Human brains have many differently functioning regions which play specialized roles in learning. By contrast, methods for training artificial neural networks, such as reinforcement-learning, typically learn exclusively via a single mechanism: gradient descent. This raises the question: might human learners’ advantage in learning efficiency over deep-learning be attributed to the interplay between multiple specialized mechanisms of learning? In this work we review a series of simulated learner systems which have been built with the aim of modeling human student’s inductive learning as they practice STEM procedural tasks. By comparison to modern deep-learning based methods which train on thousands to millions of examples to acquire passing performance capabilities, these simulated learners match human performance curves---achieving passing levels of performance within about a dozen practice opportunities. We investigate this impressive learning efficiency via an ablation analysis. Beginning with end-to-end reinforcement learning (1-mechanism), we decompose learning systems incrementally to construct the 3-mechanism inductive learning characteristic of prior simulated learners such as Sierra, SimStudent and the Apprentice Learner Architecture. Our analysis shows that learning decomposition plays a significant role in achieving data-efficient learning on par with human learners---a greater role even than simple distinctions between symbolic/subsymbolic learning. Finally we highlight how this breakdown in learning mechanisms can flexibly incorporate diverse forms of natural language and interface grounded instruction, and discuss opportunities for using these flexible learning capabilities in interactive task learning systems that learn directly from a user’s natural instruction.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31290 2024-05-20T15:07:53-07:00 Yangyang Yu yyu44@stevens.edu Haohang Li hli113@stevens.edu Zhi Chen zchen100@stevens.edu Yuechen Jiang yjiang52@stevens.edu Yang Li yli269@stevens.edu Denghui Zhang dzhang42@stevens.edu Rong Liu rliu20@stevens.edu Jordan W. Suchow jws@stevens.edu Khaldoun Khashanah kkhashan@stevens.edu

Recent advancements in Large Language Models (LLMs) have exhibited notable efficacy in question-answering (QA) tasks across diverse domains. Their prowess in integrating extensive web knowledge has fueled interest in developing LLM-based autonomous agents. While LLMs are efficient in decoding human instructions and deriving solutions by holistically processing historical inputs, transitioning to purpose-driven agents requires a supplementary rational architecture to process multi-source information, establish reasoning chains, and prioritize critical tasks. Addressing this, we introduce FinMem, a novel LLM-based agent framework devised for financial decision-making. It encompasses three core modules: Profiling, to customize the agent's characteristics; Memory, with layered message processing, to aid the agent in assimilating hierarchical financial data; and Decision-making, to convert insights gained from memories into investment decisions. Notably, FinMem's memory module aligns closely with the cognitive structure of human traders, offering robust interpretability and real-time tuning. Its adjustable cognitive span allows for the retention of critical information beyond human perceptual limits, thereby enhancing trading outcomes. This framework enables the agent to self-evolve its professional knowledge, react agilely to new investment cues, and continuously refine trading decisions in the volatile financial environment. We first compare FinMem with various algorithmic agents on a scalable real-world financial dataset, underscoring its leading trading performance in stocks. We then fine-tuned the agent's perceptual span and character setting to achieve a significantly enhanced trading performance. Collectively, FinMem presents a cutting-edge LLM agent framework for automated trading, boosting cumulative investment returns.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31291 2024-05-20T15:07:55-07:00 Bonan Zhao zbn.dale@gmail.com Natalia Vélez d9f8b378a09475911509d27c0417f28d@example.org Thomas L. Griffiths dd8816ccfd6ab36ff2756b6d2ec95fb4@example.org

Human learning does not stop at solving a single problem. Instead, we seek new challenges, define new goals, and come up with new ideas. Unlike the classic explore-exploit trade-off between known and unknown options, making new tools or generating new ideas is not about collecting data from existing unknown options, but rather about create new options out of what is currently available. We introduce a discovery game designed to study how rational agents make decisions about pursuing innovations, where discovering new ideas is a process of combining existing ideas in an open-ended compositional space. We derive optimal policies of this decision problem formalized as a Markov decision process, and compare people's behaviors to the model predictions in an online behavioral experiment. We found evidence that people both innovate rationally, guided by potential returns in this discovery game, and under- and over-explore systematically in different settings.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence https://ojs.aaai.org/index.php/AAAI-SS/article/view/31292 2024-05-20T15:07:56-07:00 Bonan Zhao zbn.dale@gmail.com Christopher G. Lucas c9daea6f7ec1abfddcbf8c853cd1bee8@example.org Neil R. Bramley 4533e1851f52654676ecd269c6995b3f@example.org

We propose bootstrap learning as a computational account for why human learning is modular and incremental, and identify key components of bootstrap learning that allow artificial systems to learn more like people. Originated from developmental psychology, bootstrap learning refers to people's ability to extend and repurpose existing knowledge to create new and more powerful ideas. We view bootstrap learning as a solution of how cognitively-bounded reasoners grasp complex environmental dynamics that are far beyond their initial capacity, by searching ‘locally’ and recursively to extend their existing knowledge. Drawing from techniques of Bayesian library learning and resource rational analysis, we propose a computational modeling framework that achieves human-like bootstrap learning performance in inductive conceptual inference. In addition, we demonstrate modeling and behavioral evidence that highlights the double-edged sword of bootstrap learning, such that people processing the same information in different batch orders could induce drastically different causal conclusions and generalizations, as a result of the different sub-concepts they construct in earlier stages of learning.

2024-05-20T00:00:00-07:00 Copyright (c) 2024 Association for the Advancement of Artificial Intelligence