Understanding trained CNNs by indexing neuron selectivity

The impressive performance and plasticity of convolutional neural networks to solve different vision problems are shadowed by their black-box nature and its consequent lack of full understanding. To reduce this gap we propose to describe the activity of individual neurons by quantifying their inherent selectivity to specific properties. Our approach is based on the definition of feature selectivity indexes that allow the ranking of neurons according to specific properties. Here we report the results of exploring selectivity indexes for: (a) an image feature (color); and (b) an image label (class membership). Our contribution is a framework to seek or classify neurons by indexing on these selectivity properties. It helps to find color selective neurons, such as a red-mushroom neuron in layer conv4 or class selective neurons such as dog-face neurons in layer conv5, and establishes a methodology to derive other selectivity properties. Indexing on neuron selectivity can statistically draw how features and classes are represented through layers at a moment when the size of trained nets is growing and automatic tools to index can be helpful.

Understanding learned CNN features through Filter Decoding with Substitution

In parallel with the success of CNNs to solve vision problems, there is a growing interest in developing methodologies to understand and visualize the internal representations of these networks. How the responses of a trained CNN encode the visual information is a fundamental question both for computer and human vision research. Image representations provided by the first convolutional layer as well as the resolution change provided by the max-polling operation are easy to understand, however, as soon as a second and further convolutional layers are added in the representation, any intuition is lost. A usual way to deal with this problem has been to define deconvolutional networks that somehow allow to explore the internal representations of the most important activations towards the image space, where deconvolution is assumed as a convolution with the transposed filter. However, this assumption is not the best approximation of an inverse convolution. In this paper we propose a new assumption based on filter substitution to reverse the encoding of a convolutional layer. This provides us with a new tool to directly visualize any CNN single neuron as a filter in the first layer, this is in terms of the image space.