Dynamic Probabilistic Pruning: A General Framework for Hardware-Constrained Pruning at Different Granularities.

Area of Science:

• Deep Learning
• Neural Network Compression
• Computer Vision

Background:

• Unstructured neural network pruning achieves high compression but results in irregular sparse matrices, hindering hardware efficiency.
• Structured pruning methods offer hardware efficiency but lack flexibility, pruning entire layers or feature maps.
• Existing methods present a trade-off between compression flexibility and hardware implementation efficiency.

Purpose of the Study:

• To introduce a flexible pruning mechanism that balances fine-grained and coarse-grained pruning.
• To enable efficient memory organization during neural network pruning.
• To facilitate joint optimization of pruning and weight quantization for enhanced network compression.

Main Methods:

• Dynamic Probabilistic Pruning (DPP) algorithm leveraging Gumbel-softmax relaxation for differentiable k-out-of-n sampling.
• Pruning at multiple granularities: weights, kernels, and feature maps.
• Joint optimization of pruning masks and weight quantization.
• Development of novel information-theoretic metrics for pruning mask analysis.

Main Results:

• DPP achieves competitive compression ratios and classification accuracy on benchmark image datasets.
• The algorithm maintains efficient memory organization, pruning specific numbers of weights or kernels.
• Joint optimization with weight quantization further enhances network compression.
• Novel metrics provide insights into pruning mask confidence and diversity.

Conclusions:

• DPP offers a flexible and efficient approach to neural network pruning, overcoming limitations of existing methods.
• The method facilitates hardware-efficient implementations through structured yet adaptable sparsity patterns.
• DPP enables synergistic compression strategies, including joint pruning and quantization.