TL;DR
This paper introduces a Generative Adversarial Inference network that learns to infer latent stimuli from noisy inputs, naturally reproducing human-like perceptual bias reversals driven by uncertainty, unifying normative Bayesian theory with deep learning.
Contribution
It presents a novel GAI network that learns inference strategies directly from sensory data, explaining perceptual bias emergence without hand-crafted priors.
Findings
The network reproduces human perceptual bias reversal under uncertainty.
Emergent behavior reflects efficient coding and Bayesian inference signatures.
Provides an end-to-end model unifying normative theory and deep learning.
Abstract
Perceptual estimates exhibit a reversal in bias depending on uncertainty: they shift toward prior expectations under high stimulus noise, but away from them when sensory noise dominates. The normative framework of a Bayesian observer model can account for this phenomenon, yet most formulations treat it as given rather than explaining its emergence through learning. We introduce a Generative Adversarial Inference (GAI) network that acquires latent representations and inference strategies directly from sensory inputs, without hand-crafted likelihoods or priors. Trained using adversarial learning with reconstruction on Gabor stimuli under varying uncertainty, the network learns to recover underlying stimuli from noisy inputs, and spontaneously reproduces the bias reversal observed in human perception. This emergent behavior arises from network responses that reveal signatures of efficient…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
The paper is very well-written - I am by no means an expert in this field, but I found the exposition clear and the experiments natural. The technique is novel to my knowledge, and I think it's quite interesting that a "fully learned" approach recovers cognitive biases actually exhibited by humans without hard-coding these biases in.
Most of my concerns are about the experimental setup of the paper: 1. The paper is restricted to very structured 2D inputs. The authors acknowledge this in the limitations section, but I didn't fully understand why extending this to color images (even CIFAR-10-sized images) would be difficult. 2. Most of the comparisons to human studies seem to be rather qualitative, it would be nice to have more quantitative results matching the two biases. 3. As far as I'm aware, the Gabor patch dataset us
The paper has the following strength: 1. Address a well-defined problem that is how to build a model to explain the attractive and repulsive perceptual bias without any hand-crafted priors or likelihood 2. The experimental results are clear and show that GAI can be trained from raw data and can replicate the attractive and repulsive phenomenon (Fig. 2, 3), and the model can reconstruct the prior under high sensory noise. The paper also includes ablation study to show the importance of having bot
1. While the problem is well-defined and the experimental results are clear and easy to understand, the scope of the experimental results may be quite narrow, since the training dataset is very simple (synthetic data of grayscale Gabor patch with bimodal priors). Since Generative Adversarial training usually has problem with scalability and mode collapsing, I'm curious whether the GAI framework can still work in the settings with larger dataset and multi-modal distribution. While I understand th
1. The paper addresses an important question in perceptual modeling. The setup is conceptually clean and connects well to long-standing debates around Bayesian and efficient-coding accounts of perception. 2. The experimental setup is clear and minimal, making it easy to interpret the results. The paper describes the pipeline, objectives, and data generation process transparently, and the figures are well thought out. 3. The paper uses useful metrics/probes (such as the Fisher information peak)
1. My biggest concern is that study is limited to 32×32 Gabor patches generated from a manually defined bimodal prior. This setup is far removed from naturalistic perception and restricts the generality of the findings. As it stands, the work feels more like a proof-of-concept demonstration than a full investigation of how perceptual biases emerge in learned systems 2. The model’s bias reversals may arise from properties of the decoding procedure rather than genuinely Bayesian inference. “Exter
This paper presents an interesting study on how the GAI can learn the perceptual-estimation biases end-to-end without any prior. It made a simple adaptation from BiGAN by adding a reconstruction noise to form GAI. Though the method is only a simple adaptation, the resulting ablation study showed that the reconstruction loss actually plays a key role in learning the biases, which is quite interesting and is what I really like about this work.
1. The paper mainly focuses on Gabor images and doesn’t show any evidence that this could be extended to other more naturalistic domains such as natural images, despite using a GAN-based framework which should be capable of these tasks. 2. disagree with the authors that, quote, “*However, these models do not account for an essential factor of perceptual processing—uncertainty in inputs and representations—and thus no existing framework unifies efficient representation with uncertainty-sensitiv
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
