Intermediate-grained kernel elements pruning with structured sparsity.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) require significant computational resources, hindering deployment on embedded or mobile devices.
• Current structured pruning methods often lead to unacceptable accuracy degradation at high pruning rates.
• There is a need for structured pruning techniques that balance high compression ratios with minimal impact on classification accuracy.

Purpose of the Study:

• To introduce Kernel Elements Pruning (KEP), a novel structured pruning method for CNNs.
• To develop a technique that achieves high pruning rates with minimal accuracy decline.
• To ensure the general applicability of structured pruning across different hardware architectures.

Main Methods:

• KEP explores the significance of individual elements within each kernel plane to identify and remove unimportant weights.
• A controllable regularization penalty is applied using a prior knowledge mask to achieve model compactness.
• A sparse convolution operation, distinct from traditional sliding windows, is introduced to eliminate redundant calculations during forward inference, optimizing for FPGA deployment.

Main Results:

• KEP demonstrates effectiveness on CIFAR-10 and ImageNet datasets.
• The method significantly reduces parameters and floating-point operations (FLOPs) compared to state-of-the-art structured methods.
• KEP maintains strong performance even with a low number of non-zero weights, indicating efficient compression.

Conclusions:

• KEP is an effective structured pruning technique for compressing deep convolutional neural networks.
• The method achieves superior parameter and FLOPs reduction while preserving classification accuracy.
• KEP's sparse convolution operation offers advantages for deployment on hardware accelerators like FPGAs.