Explore the Hierarchical Auditory Information Processing Via Deep Convolutional Autoencoder

Liting Wang, Xintao Hu, Huan Liu, Heng Huang, Lei Guo, Tianming Liu

Abstract: Combined with neural encoding models, hierarchical feature representation of sensory information via deep neural network (DNN) has been used to explore the hierarchical organization of sensory cortices. With those advancements, previous studies have revealed a representational gradient in the superior temporal gyrus (STG) in auditory information processing, where hierarchical feature representation of auditory stimuli used in fMRI experiments is derived in a supervised manner, that is, the DNN models are trained to classify auditory stimuli. However, feature representation is biased towards discriminative ones in such a supervised DNN and consequently may contaminate brain encoding models. In this study, we propose to derive hierarchical features of auditory stimuli via unsupervised DNN, namely, deep convolutional auto-encoder (DCAE), and develop an encoding model based on LASSO algorithm to explore the relationship between features in multilayers and fMRI brain responses. The results show that auditory cortex is more sensitive to low-level features represented in shallower layers whereas the visual cortex and insula are more sensitive to high-level features represented in deeper layers. These results may provide novel evidence to understand the hierarchical auditory information processing in the human brain.