Neural oscillatory activities in separate frequencies encode hierarchically distinct visual features

Although most previous studies in cognitive neuroscience have focused on the change of the neuronal firing rate under various conditions, there has been increasing evidence that indicates the importance of neuronal oscillatory activities in cognition. In the visual cortex, specific time-frequency bands are thought to have selectivity to visual stimuli. Furthermore, several recent studies have shown that several time-frequency bands are related to frequency-specific feedforward or feedback processing in inter-areal communication. However, few studies have investigated detailed visual selectivity of each time-frequency band, especially in the primate inferior temporal cortex (ITC). In this work, we analyze frequency-specific electrocorticography (ECoG) activities in the primate ITC by training encoding models that predict frequency-specific amplitude from hierarchical visual features extracted from a deep convolutional neural network (CNNs). We find that ECoG activities in two specific time-frequency bands, theta (around 5 Hz) and gamma (around 20-25 Hz) bands, are better predicted from CNN features than the other bands. Furthermore, theta- and gamma-band activities are better predicted from higher and lower layers in CNNs, respectively. Our visualization analysis using CNN-based encoding models qualitatively show that theta- and gamma-band encoding models have selectivity to higher- and lower-level visual features, respectively. Our results suggest that neuronal oscillatory activities in theta and gamma bands carry distinct information in the hierarchy of visual features, and that distinct levels of visual information are multiplexed in frequency-specific brain signals.