Local and distributed information coding in the ventral stream

Abstract

Neuroscience is replete with evidence that cognitive representations are distributed across many cortical regions. Yet, the scale and content of such distributed processing is unclear. Do findings of widespread information coding suggest a large-scale "forest" of regions interacting to represent information or instead imply a multitude of small-scale trees, processing information as localized modules. To investigate this distinction, we used visual and semantic representational analysis of fMRI data from 60 participants viewing everyday objects in multiple task contexts, and we examined the relationships between regions in terms of information coding. We demonstrate that coding of visual content in the occipital lobe is overwhelmingly modular, such that different occipital structures show limited coordination and tend to encode information redundantly. By contrast, the coding of semantic content in the inferior temporal lobe involves a high degree of coordination between regions, which optimize their coding to collectively represent a large semantic space with minimal redundancy between regions. No other brain area - neither the parietal nor prefrontal cortices - shows the preference for large-scale coding seen in the inferior temporal lobe. Taken together, these results outline a framework of how the ventral stream transitions from small-scale to large-scale coding as information progresses from visual to semantic representations. Significance statementHow does the brain convert incoming signals into usable information? Many studies have investigated this question by attempting to clarify which brain regions encode what information (e.g., V4 encodes color information). We instead aimed to shed light on the degree of coordination among information coding regions. We find that the visual-to-semantic transition as information flows anteriorly in the occipitotemporal cortex is accompanied by a shift from modular to distributed coding. That is, occipital regions encode perceptual information relatively independently with redundancy, while inferior temporal lobe regions cooperate to most efficiently represent a large space of semantic information. By leveraging ideas from information theory, our work introduces coding scale as a new dimension for understanding the architecture of information coding.