A Deep Learning Approach to Detecting Temporal Characteristics of Cortical Regions

Abstract

One view of the neocortical architecture is that every region functions based on a universal computational principle. Contrary to this, we postulated that each cortical region has its own specific algorithm and functional properties. This idea led us to hypothesize that unique temporal patterns should be associated with each region, with the functional commonalities and variances among regions reflecting in the temporal structure of their neural signals. To investigate these hypotheses, we employed deep learning to predict electrodes locations in the macaque brain using single channel ECoG signals. To do this, we first divided the brain into seven regions based on anatomical landmarks, and trained a deep learning model to predict the electrode location from the ECoG signals. Remarkably, the model achieved an average accuracy of 33.6%, significantly above the chance level of 14.3%. All seven regions exhibited above-chance prediction accuracy. The feature vectors of models identified two main clusters: one including higher visual areas and temporal cortex, and another encompassing the remaining other regions. These results bolster the argument for unique regional dynamics within the cortex, highlighting the diverse functional specializations present across cortical areas.