Hybrid seminar on Zoom and in the GERAD seminar room.

We consider the problem of training a deep neural network with non-smooth regularization to retrieve a sparse and efficient sub-structure. Our regularizer is only assumed to be lower semi-continuous and prox-bounded. We combine an adaptive quadratic regularization approach with proximal stochastic gradient principles to derive a new solver, called SR2. Our experiments on network instances trained on CIFAR-10 and CIFAR-100 with L1 and L0 regularization show that SR2 achieves higher sparsity than other proximal methods such as ProxGEN and ProxSGD with satisfactory accuracy.

Joint seminar with Alexandre Forel.