Modeling information flow through deep convolutional neural networks

In this work we investigate methods for optimizing deep convolutional neural networks (CNN) by 1) reducing the computational complexity and 2) improving classification performance for the task of transfer learning. Based on the work of Chaddad et al. (2019, 2017), the CNN is modeled as a Markov chain, where the filter output at a layer is conditionally independent of the rest of the network, given a set of previous layers. Filter banks at each layer are compressed using principal component analysis (PCA), where a reduced set of orthogonal basis filters are used to reduce the number of convolutions required while preserving classification accuracy. Information theory is then used to quantify the flow of image information through the network. Filter responses with low conditional entropy (CENT) are shown to be highly effective in image classification, and can be used as generic features for effective, noise resistant transfer learning. CENT feature analysis is demonstrated in various contexts including computer-assisted diagnosis of Alzheimer’s disease (AD) from 3D magnetic resonance images (MRI) of the human brain, and object classification in 2D photographs.