Multidimensional Pruning and Its Extension: A Unified Framework for Model Compression.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Existing model compression techniques for Convolutional Neural Networks (CNNs) often focus on single-dimension redundancy reduction (e.g., channels, spatial, or temporal).
• This limited scope can hinder optimal compression efficiency for various network architectures and data types.

Purpose of the Study:

• To propose a novel Multidimensional Pruning (MDP) framework for end-to-end compression of both 2-D and 3-D CNNs.
• To extend the MDP framework to efficiently compress Point Cloud Neural Networks (PCNNs), addressing irregular point cloud data.

Main Methods:

• The Multidimensional Pruning (MDP) framework simultaneously reduces channel redundancies and other dimensional redundancies (spatial for 2-D CNNs, spatial/temporal for 3-D CNNs).
• The MDP-Point approach extends this framework to PCNNs by targeting point dimension redundancy.
• The methods were applied and evaluated on six benchmark datasets.

Main Results:

• Demonstrated the effectiveness of the MDP framework in compressing 2-D and 3-D CNNs.
• Showcased the successful application of the extended MDP-Point approach for PCNN compression.
• Comprehensive experiments confirmed the significant compression capabilities across diverse datasets.

Conclusions:

• The proposed MDP framework offers a more comprehensive approach to model compression by considering multiple dimensions.
• MDP and its extension MDP-Point provide effective solutions for compressing various neural network architectures, including CNNs and PCNNs.
• This work advances the field of model compression, enabling more efficient deployment of deep learning models.