A Bregman Learning Framework for Sparse Neural Networks

Explore a cutting-edge learning framework for sparse neural networks in this virtual seminar talk from the 30th Imaging & Inverse Problems (IMAGINE) OneWorld SIAM-IS series. Delve into Leon Bungert's presentation on a novel approach using stochastic Bregman iterations, which enables training of sparse neural networks through an inverse scale space method. Learn about the baseline LinBreg algorithm, its accelerated momentum version, and AdaBreg, a Bregmanized generalization of the Adam algorithm. Discover a statistically sound sparse parameter initialization strategy and gain insights into stochastic convergence analysis of loss decay, along with additional convergence proofs in the convex regime. Understand how this Bregman learning framework can be applied to Neural Architecture Search, potentially uncovering autoencoder architectures for denoising or deblurring tasks.