Visualization of maximizing images with deconvolutional optimization method for neurons in deep neural networks

Deep neural networks have already proved their efficiency in solving various types of machine learning problems, especially related to recognizing natural images. However, we still dont have an exhausting understanding of how this networks work, especially in deep hidden layers. Developing methods of visualizing an information encoded in neural networks would help to reveal what kind of features hidden layers have learned and to analyze what each neuron is actually responsible for. There are two main approaches in visualizing neural networks: deconvolution and optimization. The first one is often used because of its high speed and low difficulty, but reconstructed images do not pretend to have high accuracy. The other one is quite precise: it is formulated as an optimization problem of maximizing activity of the definite neuron but takes a lot of time to converge for the deep network. We have tried to combine these two methods in order to have a possibility for the visualization with high accuracy. We used regularization based on neurons with specific activation to make images more interpretable. (C) 2018 The Authors. Published by Elsevier B.V.