ACSL: Adaptive correlation-driven sparsity learning for deep neural network compression.

Area of Science:

• Computer Vision
• Deep Learning
• Model Compression

Background:

• Deep convolutional neural networks (CNNs) require significant computational resources, limiting deployment on edge devices.
• Existing compression techniques primarily target image classification, with limited validation on pixel-level tasks like crowd counting and semantic segmentation.

Purpose of the Study:

• To develop a novel channel pruning framework for CNNs applicable to both image-level and pixel-level tasks.
• To improve the efficiency of deep learning models without compromising accuracy.

Main Methods:

• Propose an adaptive correlation-driven sparsity learning (ACSL) framework.
• Quantify channel correlations using a channel affinity matrix.
• Induce channel sparsity via an adaptive penalty strength mechanism.

Main Results:

• ACSL outperforms state-of-the-art methods on image-level classification, semantic segmentation, and dense crowd counting.
• Achieved up to 94% parameter reduction and 84% FLOPs reduction on crowd counting tasks (VGG16-Decoder and ResNet101).
• Maintained or improved performance compared to original models across all tested tasks.

Conclusions:

• The ACSL framework offers a superior approach for compressing CNNs, demonstrating effectiveness across diverse vision tasks.
• ACSL enables the deployment of accurate and efficient deep learning models on resource-constrained edge devices.