[2J4-GS-2] Machine learning: Deep reinforcement learning

[2J4-GS-2-02] Model Compression and Acceleration by Combining Reinforcement Learning and Meta-Pruning

〇Yu Kono^1,2, Motoki Omura^3,1, Tomohiro Kato¹, Yusuke Uchida¹ (1. DeNA Co., Ltd., 2. Tokyo Denki University, 3. The University of Tokyo)

Keywords:Machine learning, Reinforcement learning, Model compression, Channel Pruning

Deep learning put to practical by advanced computers, has made it possible learning complex and highly accurate models with large number of parameters. However, when inference model in a service, the communication environment, the number of requests, the devices mounted, etc have restrictions. We focused on AMC (AutoML for model compression) which compression method based on deep reinforcement learning, among structural and efficient compression methods on CNN. AMC has a problem that the estimation accuracy during the search is different from the accuracy of the model that was learned anew after the reduction with the simplified accuracy after the reduction.However, AMC does not restrict specific reinforcement learning algorithm and channel prioritization algorithm for compressing model size. Therefore, in this study, we change the reinforcement learning algorithm and combine MetaPruning which is a method that removes the channel and learns the estimated value of the weight after re-learning in advance. So that we proposed a method which more efficiently searches the accuracy based on the predicted weight as rewards.