韓松博士畢業論文（精簡總結）

 

 

 

 

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：

1. 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
2. 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
3. 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