中国科学院自动化研究所机构知识库

Nowadays,deep neural networks (DNNs) have been equipped with powerful representation capabilities. The deep convolutional neural networks (CNNs) that draw inspiration from the visual processing mechanism of the primate early visual cortex have outperformed humans on object categorization and have been found to possess many brain-like properties. Recently,Vision Transformers (ViTs) are striking paradigms of DNNs and have achieved remarkable improvements on many vision tasks compared to CNNs. It is natural to ask how are the brain-like properties of ViTs. Beyond the model paradigm,we are also interested in the effects of factors,such as model size,multimodality,and temporality,on the ability of networks to model the human visual pathway,especially when considering that existing research has been limited to the CNNs. In this paper,we systematically evaluate the brain-like properties of 30 kinds of computer vision models varying from CNNs,ViTs to their hybrids from the perspective of explaining brain activities of the human visual cortex triggered by dynamic stimuli. Experiments on two neural datasets demonstrate that neither CNN nor transformer is the optimal model paradigm for modeling the human visual pathway. ViTs reveal hierarchical correspondences to the visual pathway as CNNs do. Moreover,we find that multi-modal and temporal networks can better explain the neural activities of large parts of the visual cortex,whereas a larger model size is not a sufficient condition for bridging the gap between human vision and artificial networks. Our study sheds light on the design principles for more brain-like networks. The code is available at https://github.com/QYiZhou/LWNeuralEncoding.