Deep learning systems as complex networks
Alberto Testolin,Michele Piccolini,Samir Suweis
DOI: https://doi.org/10.1093/comnet/cnz018
IF: 1.492
2019-06-07
Journal of Complex Networks
Abstract:<span class="paragraphSection">Thanks to the availability of large scale digital datasets and massive amounts of computational power, deep learning algorithms can learn representations of data by exploiting multiple levels of abstraction. These machine-learning methods have greatly improved the state-of-the-art in many challenging cognitive tasks, such as visual object recognition, speech processing, natural language understanding and automatic translation. In particular, one class of deep learning models, known as <span>deep belief networks</span> (DBNs), can discover intricate statistical structure in large datasets in a completely unsupervised fashion, by learning a generative model of the data using Hebbian-like learning mechanisms. Although these self-organizing systems can be conveniently formalized within the framework of statistical mechanics, their internal functioning remains opaque, because their emergent dynamics cannot be solved analytically. In this article, we propose to study DBNs using techniques commonly employed in the study of complex networks, in order to gain some insights into the structural and functional properties of the computational graph resulting from the learning process.</span>
mathematics, interdisciplinary applications