MGSC Website: http://math.uci.edu/~mgsc/
Convolutional neural networks (CNNs) have been widely used in image processing. Despite the high accuracy they have achieved in some tasks like image classification, the huge cost of computational resources can be unacceptable. In this talk, we will discuss some mathematical techniques for network compression. These may include regularization, threshoulding method, structured network pruning like channel pruning and some recent approaches in Nueral Architecture Search (NAS) such as gradient based methods. Some of the topics are from our recent work on variable splitting method and differentiable channel pruning algorithm.
Fanghui Xue is a 1st year graduate student working on neural networks..
This talk was given as part of MGSC and AMS Math Graduate Student Conference.
Pizza will be served after the talk.