Evolutionary Training of Sparse Artificial Neural Networks: A Network Science Perspective
share
Through the success of deep learning, Artificial Neural Networks (ANNs) are among the most used artificial intelligence methods nowadays. ANNs have led to major breakthroughs in various domains, such as particle physics, reinforcement learning, speech recognition, computer vision, and so on. Taking inspiration from the network properties of biological neural networks (e.g. sparsity, scale-freeness), we argue that (contrary to general practice) Artificial Neural Networks (ANN), too, should not have fully-connected layers. We show how ANNs perform perfectly well with sparsely-connected layers. Following a Darwinian evolutionary approach, we propose a novel algorithm which evolves an initial random sparse topology (i.e. an Erdős-Rényi random graph) of two consecutive layers of neurons into a scale-free topology, during the ANN training process. The resulting sparse layers can safely replace the corresponding fully-connected layers. Our method allows to quadratically reduce the number of parameters in the fully conencted layers of ANNs, yielding quadratically faster computational times in both phases (i.e. training and inference), at no decrease in accuracy. We demonstrate our claims on two popular ANN types (restricted Boltzmann machine and multi-layer perceptron), on two types of tasks (supervised and unsupervised learning), and on 14 benchmark datasets. We anticipate that our approach will enable ANNs having billions of neurons and evolved topologies to be capable of handling complex real-world tasks that are intractable using state-of-the-art methods.
READ FULL TEXT