1xN Pattern for Pruning Convolutional Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Network pruning is popular for reducing convolutional neural network (CNN) complexity.
• Maintaining model accuracy and achieving CPU speedups concurrently remains a challenge.

Purpose of the Study:

• To introduce a novel 1×N pruning pattern for CNNs.
• To address the limitations of existing pruning methods in balancing accuracy and speed.

Main Methods:

• Proposing a 1×N pruning pattern that groups consecutive output kernels.
• Implementing a filter rearrangement workflow for improved accuracy and correct convolutional operations.
• Utilizing parallelized block-wise vectorized operations for efficient computation.

Main Results:

• Demonstrated efficacy on ILSVRC-2012 dataset.
• Achieved 3.0% top-1 accuracy improvement over filter pruning for MobileNet-V2 at 50% pruning rate (N=4).
• Obtained 56.04ms inference time savings on Cortex-A7 CPU compared to weight pruning.

Conclusions:

• The 1×N pruning pattern effectively reduces CNN complexity while maintaining and improving accuracy.
• The proposed method offers significant speedups on general CPUs, overcoming previous limitations.
• The approach provides a practical solution for efficient CNN deployment.