Evolutionary HyperNet-InfoGAN Baselines for Controllable Synthetic Image Generation

Abstract

Synthetic medical data generation requires human-auditable control. It is important that latent factors are able to separate data-specific structures so clinicians can inspect, trust, and responsibly reuse synthetic samples. Building on prior evolutionary generator work, an Evolutionary HyperNet-InfoGAN is developed in which Policy Gradients with Parameter-Based Exploration (PGPE) optimizes a compact HyperNetwork that emits the weights of a larger image generator, while a discriminator/Qnetwork is trained with backpropagation. This process creates a non-stationary adversarial setting in which evolutionary search must continually adapt, but in a far smaller and more structured parameter space than direct evolution of the full generator. This model is used to evaluate a shared baseline on the MNIST dataset (28x28 grayscale, 10 classes) and then on a grayscale-converted BloodMNIST dataset (28x28, 8 classes) with a shared-z protocol in which groups of samples share continuous noise while discrete codes vary. This setup enables the direct measurement of feature-space code separation (rsense) and within-code variation (rintra). When compared to a previous PGPE baseline that searched directly over a generator with more than 500k parameters, the HyperNetwork formulation yielded cleaner code allocation, substantially more stable late-stage training, and the ability to resolve early duplicate modes to distinct digits for better mode coverage. When applied to MNIST, the baseline approached full digit coverage with fewer duplicate-code failures than prior direct search. For BloodMNIST, the preliminary baseline results show partial disentanglement and class-consistent morphology, suggesting that HyperNetworks provide a practical intermediate step toward auditable, controllable synthetic medical image generation. These baselines motivate the next stage of this work that will employ a Cultural Algorithms (CA) as a meta-level controller in order to support adaptive fitness weighting and secondary search guidance within the PGPE framework.