A CNN-Based Computational Encoding Model for Human V2 Cortex

Abstract

The computation encoding models, used to predict human brain activity from natural image stimuli, can be performed as a function simulator of human vision information process. In the traditional computational encoding models for human V2 cortex, due to the lack of higher visual feature and information processing hierarchy, it is difficult to achieve expected predict performance. Here, activated by the properties of CNN, we trained a CNN as an encoding model for human V2 cortex, which can be trained for predicting stimuli-evoked response measured by functional magnetic resonance imaging. The results reveal that the CNN-based encoding model can achieve a higher performance, proves that CNN have advantages in encoding higher visual areas. This finding provides a new framework for the human vision encoding models and helps to further understand of the human vision mechanism from the computational point view.