論文由三部分構成,也是韓松在博士期間的工作,相關論文與解析見下面:
- Deep Compression: Compressing Deep Neural Networks with Pruning, Trained Quantization and Huffman coding。 從軟件端極大的壓縮網絡的權重, 參考:https://blog.csdn.net/weixin_36474809/article/details/80643784
- DSD:Dense-sparse-dense training for deep neural networks。 一種由密集->稀疏->密集的新的網絡訓練方式,能從訓練層面提升網絡準確率, 參考:https://blog.csdn.net/weixin_36474809/article/details/85322584
- EIE:Efficient Inference Engine on Compressed Deep Neural Network . 在Deep compression的基礎上,EIE是基於硬件的稀疏網絡加速實現,硬件上達到很好的效果 。 參考:https://blog.csdn.net/weixin_36474809/article/details/85326634
動機Why:Neural networks are difficult to deploy on embedded systems with limited hardware resources.
* Computationally intensive
* Memory intensive
做法How:Co-designed the algorithm and hardware for deep learning.
* Simplify and compress DNN models
* Deaigned customized hardware for the compressed model
結果Result: Outpreforms CPU,GPU and mobile GPU by factors of 189,13,and 307
Consumes 24000,3400, and 2700 less energy than CPU,GPU and mobile GPU.
詳細內容參考:https://blog.csdn.net/weixin_36474809/article/details/85613013