- 43 views
- 65 downloads
Class Conditional Filter Pruning Based on Clustering of The Ranks of Feature Maps
-
- Author / Creator
- Razavi, Soroush
-
The success of deep learning is partly due to the sheer size of modern models. However, such large models strain the capabilities of mobile or resourceconstrained devices. Ergo, reducing the resource demands of AI models is essential before AI can be deployed on such devices. One promising solution to this challenge is filter pruning, which aims to streamline models without sacrificing performance. We propose a new approach to filter pruning that extends feature-map ranking to consider the class-by-class rank of each map. These feature vectors are clustered, and automated decision rules select the filters to be pruned. This makes our work one of the few pruning methods that is fully automated and does not require any human labor from end to end. Experiments using VGG and ResNet networks on the CIFAR-10 and ImageNet datasets show that our pruned models are more accurate than the well-known HRank algorithm and perform similarly on the CIFAR-100 dataset. In addition, we see significant reductions in power usage of models post-pruning. Finally, ethical pruning requires that an algorithm does not favor some group of data over others. We have verified that our algorithm follows the ethical pruning approach.
-
- Graduation date
- Fall 2024
-
- Type of Item
- Thesis
-
- Degree
- Master of Science
-
- License
- This thesis is made available by the University of Alberta Library with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.