A Multilayer Network-Based Approach to Represent, Explore and Handle Convolutional Neural Networks

Deep learning techniques and tools have experienced enormous growth and widespread diffusion in recent years. Among the areas where deep learning has become more widespread there are computational biology and cognitive neuroscience. At the same time, the need for tools able to explore, understand, and possibly manipulate, a deep learning model has strongly emerged. We propose an approach to map a deep learning model into a multilayer network. Our approach is tailored to Convolutional Neural Networks (CNN), but can be easily extended to other architectures. In order to show how our mapping approach enables the exploration and management of deep learning networks, we illustrate a technique for compressing a CNN. It detects whether there are convolutional layers that can be pruned without losing too much information and, in the affirmative case, returns a new CNN obtained from the original one by pruning such layers. We prove the effectiveness of the multilayer mapping approach and the corresponding compression algorithm on the VGG16 network and two benchmark datasets, namely MNIST, and CALTECH-101. In the former case, we obtain a 0.56% increase in accuracy, precision, and recall, and a 21.43% decrease in mean epoch time. In the latter case, we obtain an 11.09% increase in accuracy, 22.27% increase in precision, 38.66% increase in recall, and 47.22% decrease in mean epoch time. Finally, we compare our multilayer mapping approach with a similar one based on single layers and show the effectiveness of the former. We show that a multilayer network-based approach is able to capture and represent the complexity of a CNN. Furthermore, it allows several manipulations on it. An extensive experimental analysis described in the paper demonstrates the suitability of our approach and the goodness of its performance.