Research on a new paradigm for neural network compression

神经网络压缩新范例研究

100000

Neural network compression is crucial for deploying large models on resource-constrained devices. While existing techniques like quantization, pruning, and distillation have progressed, they are limited by high-performance loss as compression rates decrease, restricted applications, and demanding computing resources.
This research aims to overcome these challenges by introducing a novel compression paradigm. We aim to achieve significant size reductions with minimal performance impact by targeting semi-structured data representations between compressed and redundant data.
Preliminary theoretical analysis suggests that some neural networks can be compressed below 10% of their original size while maintaining low-performance loss. Furthermore, this study explores the compatibility of our proposed approach with state-of-the-art methods, paving the way for practical applications in Computer vision and Natural Language Processing.