Predictive Coding dynamics enhance model-brain similarity

Poster Session F - Tuesday, April 1, 2025, 8:00 – 10:00 am EDT, Back Bay Ballroom/Republic Ballroom

Bhavin Choksi¹ (choksi@uni-frankfurt.de), Manshan Guo^1,2, Sari Saba-Sadiya¹, Radoslaw Cichy², Gemma Roig¹; ¹Goethe University, Frankfurt, ²Freie Universität, Berlin

Predictive coding is a popular theory in neuroscience that posits that the brain, instead of passively encoding, actively predicts the world around it. Recent advances in deep learning models has led to an increased interest in integrating predictive coding inspired dynamics into artificial neural networks. Various implementations of these recurrent dynamics have been demonstrated to induce human-like properties, such as robustness to noise and ability to perceive illusions, into the neural networks. While these brain-inspired architectural biases have led to interesting properties, it remains unclear if they can improve the alignment between the brain and artificial representations. In this study, we systematically investigate the conditions under which brain-inspired modifications in predictive processing improve alignment between model and neural representations. This is achieved by employing models from the predify library that incorporates generative feedback on top of feedforward networks to perform recurrent dynamics over multiple time-steps. We investigate the representations across various layers in these networks, and their alignment with those obtained from two large neural datasets – one fMRI (Natural Scenes Dataset) and EEG (THINGSEEG). Our preliminary results suggest that, across both the datasets, progressively increasing feedback significantly increases similarity between model representations and those found in the higher-level visual brain areas. When the images are further divided based on their complexities, we found that the generative feedback especially helps when processing visual scenes with higher complexity, in accordance with the currently assumed role of feedback connections in the brain for processing difficult stimuli.

Topic Area: PERCEPTION & ACTION: Vision