Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Deep predictive coding networks partly capture neural signatures of short-term temporal adaptation in human visual cortex

Amber Marijn Brands¹, Paulo Ortiz², Iris Groen¹; ¹University of Amsterdam, ²Donders Institute for Brain, Cognition and Behaviour

Presenter: Amber Marijn Brands

Predictive coding is a prominent theory of cortical function that proposes that the brain continuously generates predictions about sensory inputs through a hierarchical network of top-down and bottom-up connections. Prior studies demonstrate that PredNet, a deep neural network built on predictive coding principles, indeed captures key characteristics of neural responses observed in primate visual cortex. However, one widespread neural phenomenon that remains unexplored in this context is short-term visual adaptation: the modulation of neural activity over time in response to static visual inputs that are prolonged or repeated. Here, we investigate whether PredNet exhibits two hallmark signatures of temporal adaptation previously identified in human intracranial recordings. We find that, similar to human visual cortex, activations of error units in the first layer of PredNet exhibit subadditive temporal summation to prolonged stimuli, reflecting nonlinear accumulation of response magnitude with increased stimulus duration. However, unlike the neural data, PredNet shows systematic responses to stimulus offsets. For repeated stimuli, PredNet exhibits slight response suppression for consecutively presented images, but no repetition suppression, a stronger response reduction to identical than non-identical images that is robustly observed in visual cortex. These discrepancies are consistent across different training diets, optimization strategies and model unit types. Overall, our results show that PredNet's activation dynamics only partly capture short-term temporal adaptation signatures in human visual cortex, suggesting that this particular instantiation of predictive coding does not fully account for neural adaptation phenomena.

Topic Area: Visual Processing & Computational Vision

proceeding: Full Text on OpenReview