Michail Loulakis

Michail Loulakis, Charalambos G. Makridakis

In this work we consider a model problem of deep neural learning, namely the learning of a given function when it is assumed that we have access to its point values on a finite set of points. The deep neural network interpolant is the the resulting approximation of f, which is obtained by a typical machine learning algorithm involving a given DNN architecture and an optimisation step, which is assumed to be solved exactly. These are among the simplest regression algorithms based on neural networks. In this work we introduce a new approach to the estimation of the (generalisation) error and to convergence. Our results include (i) estimates of the error without any structural assumption on the neural networks and under mild regularity assumptions on the learning function f (ii) convergence of the approximations to the target function f by only requiring that the neural network spaces have appropriate approximation capability.

Eleni Stai, Michail Loulakis, Symeon Papavassiliou

Network Utility Maximization (NUM) is often applied for the cross-layer design of wireless networks considering known wireless channels. However, realistic wireless channel capacities are stochastic bearing time-varying statistics, necessitating the redesign and solution of NUM problems to capture such effects. Based on NUM theory we develop a framework for scheduling, routing, congestion and power control in wireless multihop networks that considers stochastic Long or Short Term Fading wireless channels. Specifically, the wireless channel is modeled via stochastic differential equations alleviating several assumptions that exist in state-of-the-art channel modeling within the NUM framework such as the finite number of states or the stationarity. Our consideration of wireless channel modeling leads to a NUM problem formulation that accommodates non-convex and time-varying utilities. We consider both cases of non orthogonal and orthogonal access of users to the medium. In the first case, scheduling is performed via power control, while the latter separates scheduling and power control and the role of power control is to further increase users' optimal utility by exploiting random reductions of the stochastic channel power loss while also considering energy efficiency. Finally, numerical results evaluate the performance and operation of the proposed approach and study the impact of several involved parameters on convergence.