Filters in Convolutional Neural Networks as Independent Detectors of Visual Concepts

With the development of Convolutional neural networks (CNN), a myth was formed in recent years that only the last convolutional layers, including the fully connected layers, provide the most essential visual features (visual concepts) of the images at the input. In this paper, we experimentally disprove this prejudice through demonstrations showing that all the convolution layers of a CNN contain essential information that influences the classification. In connection with the experiments conducted, we look at each filter from different convolution layers as an independent detector of a particular visual structure (visual concept), and on this basis we introduce a new notion called the "vector of visual concepts". This concept summarizes the information, what visual constellation detectors are activated by an input image, and how strong they are activated. The classification error obtained by the vector of visual concepts differs by only 0.12% from the results of the last (softmax) layer of the network on the training data domain, which undoubtedly suggests the information saturation of all layers in the network. Additionally, we show by experiments that in input data from a domain other than the training one, the features of the initial and intermediate convolution layers provide a more accurate classification than those of the last convolution layers.