Deep Learning Seminar  /  February 25, 2021, 10:00 – 11:00

From Network Pruning to Network Architecture Search: A Unified Framework based on Group Sparsity

Speaker: Avraam Chatzimichailidis (Fraunhofer ITWM, Department High Performance Computing)

Abstract:

Network Pruning aims at reducing the neural network size with an acceptable egradation in the performance. Depending on the granularity, there are weight pruning, kernel pruning and filter pruning. The different pruning objectives are treated in literature separately, and various methods are proposed for each pruning objective. Another important topic in deep learning is network architecture search, which aims at finding a good architecture for the given problem at hand. In this seminar, we show how these seemingly unrelated problems can be effectively tackled under a unified framework based on group sparsity.