TL;DR
처음으로 Neural Network compression 제안.
- pruning
- 9~13X
- quantization
- 32bits to 5bits
- huffman coding
→ accuracy 영향없이 storage saving을 30~40X 늘림.
privacy, network bandwidth, storage, energy consumption 이점이 있음.
Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman Coding
Category
Network Pruning
Network Quantization
Year/Month
2015-10
Status
Publications
ICLR
Code