Sparsification and Sparsity-Preserving PSD Correction for Cross-Layer Aware Mixed-Precision Quantization

Key Points

• This research aims to address the limitations of existing mixed-precision quantization methods in deep learning by preserving sparsity and ensuring positive semidefinite corrections.
• Proposed a new framework incorporating sparsification and PSD correction techniques.
• Reflected on the hierarchical structure of deep learning models to reduce unnecessary interactions.
• Evaluated using ResNet-34 and ImageNet-1K dataset for optimization assessment.
• Achieved higher Top-1 accuracy with fewer computations compared to existing methods.
• Demonstrated effective preservation of matrix sparsity during optimization.