Attractive and Repulsive Perceptual Biases Naturally Emerge in Generative Adversarial Inference

Human perceptual estimates exhibit a striking reversal in bias depending on uncertainty: they shift toward prior expectations under high sensory uncertainty, but away from them when internal noise is dominant. While Bayesian inference combined with efficient coding can explain this dual bias, existing models rely on handcrafted priors or fixed encoders, offering no account of how such representations and inferences could emerge through learning. We introduce a Generative Adversarial Inference (GAI) network that simultaneously learns sensory representations and inference strategies directly from data, without assuming explicit likelihoods or priors. Through joint reconstruction and adversarial training, the model learns a representation that approximates an efficient code consistent with information-theoretic predictions. Trained on Gabor stimuli with varying signal-to-noise ratios, GAI spontaneously reproduces the full transition from prior attraction to repulsion, and recovers the Fisher information profile predicted by efficient coding theory. It also captures the characteristic bias reversal observed in human perception more robustly than supervised or variational alternatives. These results show that a single adversarially trained network can jointly acquire an efficient sensory code and support Bayesian-consistent behavior, providing a neurally plausible, end-to-end account of perceptual bias that unifies normative theory and deep learning.