Visualizing And Understanding Deep Neural Networks

Deep neural networks (DNNs) have achieved tremendous progress on many pat-tern recognition tasks, especially the large-scale images recognition problem. But DNNs still make mistakes easily. And it is still unclear how DNNs learn suitable features from the training data. The drawback and "black-box" property of DNNs promote the de-velopment of deep visualization. Many works have concentrated on visualizing and un-derstanding the inner mechanism of deep neural networks by generating an image that activates some specific neurons. Deep visualization aims to provide some insights that help researchers understand how DNNs learn feature from big data, analyze the merit and demerit of DNNs and improve performance with a better DNNs.Deep visualization has attracted many researchers’ attention, but it is just a begin-ning. This paper aims to visualize and understand the thousands of convolutional filters in DNNs. The features extracted by convolutional filters in DNNs are diverse. The filters at low layers capture low semantic representation, such as edge and corner while the filters at high layers capture high semantic representation, such as image category. However, it is still visually unclear what every filter extracts from input images.In this paper, we propose a modified code inversion algorithm, called feature map inversion, to understand the function of the filter of interest in DNNs. By enhancing the assigned feature map and weakening the others, the visualizations of feature map inver-sion reveal that every filter in DNNs is trying to capture a specific texture primitive. The texture primitives for low layers generate images whose color is monotonous and the lo-cal structure is simple. As layers increase higher, the colours become plentiful and the local structures become more intricate. Based on these understandings, we propose a simple texture representation and apply it to image style transfer and texture recognition task. Our texture representation can not only be used to generate qualitative images of di-verse styles, but also achieve outstanding performance on texture recognition. We expect our work could help understand how DNNs extract high semantic information from input images and provide some helpful insights for designing more effective architectures.