Network pruning for visual food recognition

Machine learning has become very popular in recent years due to its great learning ability that can be applied to various tasks. Convolutional Neural Network (CNN) is one of the most popular types in machine learning and many advanced CNN architectures are designed. However, most of these networks a...

Full description

Saved in:
Bibliographic Details
Main Author: Yao, Yujie
Other Authors: Yap Kim Hui
Format: Thesis-Master by Coursework
Language:English
Published: Nanyang Technological University 2020
Subjects:
Online Access:https://hdl.handle.net/10356/141417
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Machine learning has become very popular in recent years due to its great learning ability that can be applied to various tasks. Convolutional Neural Network (CNN) is one of the most popular types in machine learning and many advanced CNN architectures are designed. However, most of these networks are made up of a large number of parameters which can’t be easily applied to mobile applications. Therefore, networks with fewer parameters are in demand. Network Pruning and compact network are two popular methods to reduce the size of networks. Network Pruning is a method which can remove unimportant weights in the network and thus, reducing the number of parameters and computation complexity. There are different algorithms in network pruning such as iterative pruning and inference-time pruning. In this thesis, we mainly focus on one pruning algorithm which is the lottery ticket hypothesis. This method states that dense neural networks contain much smaller subnetworks which can reach the similar accuracy to the original network based on their results. In our project, we first apply this pruning method to several CNN architectures in smaller dataset, MNIST, to test its performance. Then we train deeper CNNs in larger datasets such as CIFAR-10 and CIFAR-100 to further evaluate the results. When the previous results are all reasonable, we apply the pruning algorithm in UEC-FOOD256 dataset by training two CNNs. In our experimental results, the accuracies obtained in food dataset are better than the results in CIFAR-100 that the accuracies only drop by less than 3% with nearly 50% of weights in original network being pruned.