Layer Dependent Artificial Representation and Selectivity of Model Neurons in the AlexNet Model Trained for Object Classification

The analysis and understanding of a trained deep convolutional neural network (DCNN) model facilitates our understanding of not only the artificial approach for explaining the information processing involved in object classification but also the neural mechanism used to establish the visual perception of an object. Previous studies found that an analysis of trained DCNN models, like the AlexNet model, using stylized images, derived from natural images by altering their texture while preserving their fundamental contours and object shapes, provides insight into the characteristics of artificial representation in hidden layers for object classification. Additionally, the classification of objects in stylized images by trained DCNN models tends to be distinct from that on the original images. However, the detailed characteristics of selectivity in model neurons for classifying objects are still not fully understood at the layer level of a trained AlexNet model. In this study, we investigated the characteristics of the artificial representations and selectivity in each layer of a trained AlexNet model by computing t-SNE embeddings of the responses of model neurons to stylized images generated using various natural images and natural object surface images. We found that these embeddings for the early- and intermediate-level layers showed a clear clustering based on fundamental visual features in natural images. By contrast, the responses of model neurons in fully connected layers were clustered according to the textures and styles arising from the natural object surface images used for generating stylized images. These results suggest not only the layer-specific selectivity of model neurons in the trained AlexNet model but also the correspondences and differences between the AlexNet model’s artificial approach to classifying objects and neural mechanisms involved in object perception.