Convolutional Neural Networks with Sparse Connections along the Depth Dimension

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

Final Year Thesis Oral Defense

Title: "Convolutional Neural Networks with Sparse Connections along the 
        Depth Dimension"

by

WONG Ngo Yin

Abstract:

Deep convolutional neural networks demand high computational and memory 
resources, which is difficult for them to be deployed on systems with 
limited resources. Network pruning techniques are widely used to prune the 
weights and filters of a deep convolutional neural network to reduce their 
cost. In this final year thesis, we propose a method to prune a trained 
network based on the structure of the training data. Each layer of the 
network is rebuilt by analyzing the structure of the output of the 
previous layer. More specifically, we learn a tree-structured 
probabilistic model from the output using Chow-Liu's algorithm, analyze 
strongly correlated features and prune the unimportant connections. The 
resulting model is more compact.


Date            : 3 May 2019 (Friday)

Time            : 11:10 - 11:50

Venue           : Room 5566 (near lifts 27/28), HKUST

Advisor         : Prof. ZHANG Nevin Lianwen

2nd Reader      : Prof. YEUNG Dit-Yan