Using Structured Pruning to Find Winning Lottery Tickets

Abstract

In recent years, deep neural networks have successfully solved artificial intelligence problems. However, large models need more memory and computational resources. Recent research has proved that the deep models are over-parameterized and have redundancy in their parameterization. The lottery ticket hypothesis paper by Frankle and Carbin offers that based on pruning, we can achieve subnetworks with initializations that are capable of training from scratch. Still, they have used unstructured pruning, and the resulting architectures are sparse that need special hardware/software for compression and speedup. On the other hand, structured pruning methods in the convolutional neural networks (CNNs) preserved the structure of the convolution layers. Therefore, we do not need special hardware/software (HW/SW) libraries. In this work, we examined the lottery ticket hypothesis with structured pruning techniques and used these methods with different architectures.