Anish Desai

 

Date:  Tuesday, April 16, 2019

Time:  1:30 pm – 3:30 pm

Location: 3105, Lambton Tower

The success of Deep Neural Networks (DNN) in classification is accompanied by a drastic increase in weight parameters which also increases the computational and storage costs. Pruning of DNN involves identifying and removing redundant parameters with little or no loss of accuracy. Layer-wise pruning of weights by their magnitude has shown to be an efficient method to prune neural networks. However, finding the optimal values of threshold for each layer is a challenging task given the large search space. To solve this problem, we use multi-population cultural algorithm which is an evolutionary algorithm that takes advantage of knowledge domains and faster convergence and is used in many optimization problems. We experiment it on LeNet-style models and measure level of pruning through the pruning ratio. Results show that our method achieves the best pruning ratio compared with some state-of-the-art DNN pruning methods. By removing redundant parameters, the computational and storage costs are reduced significantly. 

 

Thesis Committee:

Thesis Defense Announcement

 

5113 Lambton Tower, 401 Sunset Ave., Windsor ON, N9B 3P4 (519) 253-3000 Ext 3716, csgradinfo@uwindsor.ca