SOKS: Automatic Searching of the Optimal Kernel Shapes for Stripe-Wise Network Pruning

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep convolutional neural networks (CNNs) exhibit high performance but suffer from over-parameterization and computational expense.
• Existing network pruning methods often lack structural constraints or overlook kernel shape redundancy.

Purpose of the Study:

• To develop a framework for automatically searching optimal kernel shapes and performing efficient network pruning.
• To address the redundancy within convolution kernels and their impact on CNN performance.

Main Methods:

• Introduced a framework termed Searching of the Optimal Kernel Shape (SOKS).
• Utilized coefficient matrices regularized by various terms to identify important kernel positions.
• Implemented stripe-wise pruning (SWP) using optimized irregular kernels.

Main Results:

• SOKS effectively searches for high-efficiency kernel shapes, enhancing CNN performance.
• Achieved superior results in model compression ratio and reduced inference latency.
• Embedding SOKS-optimized kernels into VGG-16 improved CIFAR-10 accuracy to 94.26%, pruned 59.27% parameters, and reduced latency by 27.07%.

Conclusions:

• SOKS offers an effective approach to optimize kernel shapes for efficient CNNs.
• The method successfully reduces model size and speeds up inference without compromising accuracy.