Resource aware, highly-scalable deep learning

An industry has grown around the deployment of deep learning models on edge and embedded devices, largely based on network compression and specialized training strategies. This research aims to substantially advance the state-of-the-art in resource-efficient deep learning, both in training and deployment, by combining new ideas in discrete and model-based optimization. In particular, we will explore strategies for resource-aware risk minimization, and combine them with post-training network compression in order to optimize the tradeoff between optimization tractability and optimality of the resulting solution. Special focus will be devoted to constrained optimization settings in which specific target hardware proles are available.