Layer-Specific Characteristics of Artificial Representations in the Trained AlexNet Model

Analysis of trained deep convolutional neural networks (DCNNs) provides insight into the mechanism of biological visual systems for visually recognizing objects. However, the detailed characteristics of artificial representations for object classification remain unclear at the layer level of DCNNs. Previous studies have found that stylized images generated from natural images by changing their textures may underlie the object classification performed by trained DCNN models (such as the AlexNet model) and modulate the artificial representations of their model neurons in hidden layers. In this study, we analyzed the characteristics of the artificial representations in each layer of a trained AlexNet model by computing the correlation between natural and stylized images according to the activities of model neurons. The activities of model neurons in response to these natural and stylized images were highly correlated in the low-level layers of the trained AlexNet model. However, the magnitude of the correlation decreased as the level of the model layer increased. These results imply that a trained AlexNet model may establish texture-based representations of images through the hierarchy of its network. This finding is in agreement with results of previous studies, and implies that artificial representations for object classification performed by DCNN models correspond, at least in part, to neural representations for visual perception of objects.