Machine learning techniques, especially “deep learning” using multilayer neural networks, have been highly successful in addressing a wide variety of computational problems in many domains of science and technology. However, a detailed theoretical understanding of why and how they work is not yet available. Some of the concepts and techniques of statistical physics, in which one routinely deals with theoretical descriptions of collective properties of systems consisting of very large numbers of interacting variables, are expected to be useful in analysing how deep networks function. The inherent variability of real-world data sets can be handled by techniques developed in the statistical physics of disordered systems such as spin glasses. Also, it has been suggested that tools in statistical physics, such as renormalization-group transformations and quantum annealing, may be useful in improving the performance of deep networks. In addition, machine learning methods are now being used ...
Description