Evidence for predictive computations in a brain hierarchy during a visual search task

Abstract

Many lines of evidence suggest that the cortex functions fundamentally as a predictive system, compressing high-dimensional sensory inputs into low-dimensional representations that support internal models of the environment. These models generate top-down signals that shape and filter incoming sensory information, prioritizing unpredicted, and thus more informative, inputs for further processing. Multiple, mutually compatible computational frameworks have been proposed to explain these mechanisms, including predictive coding, predictive routing, and autoencoder-based approaches. However, a key challenge remains in empirically disentangling their relative contributions. Here, we directly compared three candidate algorithms, predictive coding, routing and autoencoders using LFP data from a brain hierarchy acquired during a visual search task. Notably, predictive coding differs from alternative models in its capacity to dynamically optimize signal passing across brain hierarchies in an input-specific manner, while the other algorithms do not consider constraints on feedback or feedforward inputs. It also goes beyond pairwise interactions, by predicting information flow in triplets of brain areas; here, V4, 7A and PFC. We considered the characteristic patterns of neural dynamics and within-area coupling that distinct algorithms produce. These were subsequently used to assess the relative evidence of each algorithm in the face of the LFP data. Our results support a hybrid account: hierarchical message passing consistent with predictive coding appears necessary to explain deep-layer activity, while predictive suppression mechanisms, aligned with predictive routing, account for superficial-layer dynamics without requiring explicit error computations.