Nader Ganaba

Nader Ganaba

As it stands, a robust mathematical framework to analyse and study various topics in deep learning is yet to come to the fore. Nonetheless, viewing deep learning as a dynamical system allows the use of established theories to investigate the behaviour of deep neural networks. In order to study the stability of the training process, in this article, we formulate training of deep neural networks as a hydrodynamics system, which has a multisymplectic structure. For that, the deep neural network is modelled using a stochastic differential equation and, thereby, mean-field type control is used to train it. The necessary conditions for optimality of the mean--field type control reduce to a system of Euler-Poincare equations, which has the a similar geometric structure to that of compressible fluids. The mean-field type control is solved numerically using a multisymplectic numerical scheme that takes advantage of the underlying geometry. Moreover, the numerical scheme, yields an approximated solution which is also an exact solution of a hydrodynamics system with a multisymplectic structure and it can be analysed using backward error analysis. Further, nonlinear stability yields the condition for selecting the number of hidden layers and the number of nodes per layer, that makes the training stable while approximating the solution of a residual neural network with a number of hidden layers approaching infinity.

Massimiliano Esposito, Nader Ganaba

As the deep neural networks are being applied to complex tasks, the size of the networks and architecture increases and their topology becomes more complicated too. At the same time, training becomes slow and at some instances inefficient. This motivated the introduction of various normalization techniques such as Batch Normalization and Layer Normalization. The aforementioned normalization methods use arithmetic operations to compute an approximation statistics (mainly the first and second moments) of the layer's data and use it to normalize it. The aforementioned methods use plain Monte Carlo method to approximate the statistics and such method fails when approximating the statistics whose distribution is complex. Here, we propose an approach that uses weighted sum, implemented using depth-wise convolutional neural networks, to not only approximate the statistics, but to learn the coefficients of the sum.