\(p\)-adic statistical field theory and deep belief networks (Q2685103): Difference between revisions

There is an investigation on the correspondence between Neural Networks (NN) and Statistical Field Theory (STF). After a short review of important achievements supporting the idea of the existence of a correspondence between different forms of the neural network architectures and quantum field theories, one starts with the presentation of basic aspects of the p-adic analysis. One defines the field of \(p\)-adic numbers, \(\mathbb{Q}_p\), its topology and the Bruhat-Schwartz space. The third section of the article is devoted to the introduction of a class of non-Archimedean statistical field theories. One defines the energy-functional \(E (v, h)\) which controls the size of thermal fluctuations performed by the field \(\{v,h\}\), its discretized form, the standard restricted Boltzmann machines (RBM) and the \(p\)-adic discrete deep belief networks (DBN). In the fourth section one proves that the \(p\)-adic DBNs are universal approximators. A discussion on obtained results and open problems is to be found in the fifth section, followed by Conclusions in the last section. References contain an impressive list of one hundred titles.