Coarse-Grained Pruning of Neural Network Models Based on Blocky Sparse Structure.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Science

Background:

• Deep neural networks (DNNs) demonstrate high performance across various fields.
• Over-parameterization in DNNs leads to significant computational demands and resource consumption.
• Efficient computation of large weight matrices is a critical challenge in deploying DNNs.

Purpose of the Study:

• To propose a novel block-based division and coarse-grained block pruning strategy for DNNs.
• To simplify and compress the fully connected structure of deep learning models.
• To accelerate weight matrix calculations through optimized storage and computation.

Main Methods:

• Weight matrices are divided into square sub-blocks using spatial aggregation.
• A coarse-grained block pruning strategy is applied to reduce model parameters.
• Pruned weight matrices are stored in the Block Sparse Row (BSR) format for efficient computation.

Main Results:

• The proposed method effectively compresses deep neural networks.
• Computational costs are reduced without substantial degradation in classification accuracy.
• The block pruning strategy demonstrates compatibility with convolutional neural networks.

Conclusions:

• The coarse-grained block pruning method offers an effective way to optimize DNNs.
• BSR storage format enhances the efficiency of computations involving pruned weight matrices.
• This approach provides a viable solution for developing more efficient and deployable deep learning models.