Network science for optimizing artificial neural networks on computer vision

Abstract

Computer Vision is one of the most diffused areas within Machine Learning and where most of the deep Artificial Neural Networks (ANNs) have emerged. Although these models achieve outstanding results on large vision tasks, their ever-increasing complexity hampers a better understanding of their internal functioning. Consequently, most methods for architecture design and optimization usually are based on randomness and trial-and-error searches. Therefore, we propose to use Network Science for improving ANNs on Computer Vision. On most of these models, neurons and synapses can be directly modeled as Complex Networks (CN). We compute several topological characteristics based on neuronal centrality from this graph representation then a clustering technique highlights the most relevant neuron types. The distribution of these neuron types represents the overall structure of the ANN, which we analyze in correlation to the model's accuracy. Our methodology allows establishing a link between topology and performance for developing mechanisms to optimize them, where we will consider architecture engineering (construction/initialization) and optimization (compression/adaptation). One of the benefits of this approach is that the obtained models are more theoretically plausible within Network Science. We will consider this approach on different neural architectures and explore their applicability for specific and small-scale tasks such as texture analysis. The developed methods will also be analyzed in multidisciplinary applications from research collaborations at the host university. The development of applied methods is a crucial step of the Ph.D., and the stay in a top-international center will also strengthen the collaboration network of our research group. (AU)