HC2022.KAIST.DonghyeonHan.v02.pdf

《HC2022.KAIST.DonghyeonHan.v02.pdf》由会员分享，可在线阅读，更多相关《HC2022.KAIST.DonghyeonHan.v02.pdf（18页珍藏版）》请在三个皮匠报告上搜索。

1、1 of 22HOTCHIPS 2022HNPU-V2:A 46.6 FPS DNN Training Processor for Real-World Environmental Adaptation based Robust Object Detection on Mobile DevicesHNPU-V2:A 46.6 FPS DNN Training Processor for Real-World Environmental Adaptation based Robust Object Detection on Mobile DevicesDonghyeon Han,Dongseok

2、 Im,Gwangtae Park,Youngwoo Kim,Seokchan Song,Juhyoung Lee,and Hoi-Jun YooSemiconductor System Lab.School of EE,KAIST2 of 22HOTCHIPS 2022HNPU-V2:A 46.6 FPS DNN Training Processor for Real-World Environmental Adaptation based Robust Object Detection on Mobile DevicesDevelopment of DNN for Mobile Platf

3、orms Smarter DNNs:#of Parameter Lightweight DNNs for Mobile Devices Quantization,weight pruning,pointwise or depthwise CONV 3)Mobile DNN Arch.2)Weight PruningZeroWNonzeroWFP32 IAQuantized IAQuantized WFP32 W1)Quantization3 of 22HOTCHIPS 2022HNPU-V2:A 46.6 FPS DNN Training Processor for Real-World En

4、vironmental Adaptation based Robust Object Detection on Mobile DevicesDisadvantages of Mobile-oriented DNNs Low Detection Accuracy in Practice Performance Degradation After Unexpected Situations Low network capacity Loosing generality Sensitive to accidentAccuracy Efficiency Speed Accuracy Efficienc

5、y Speed Low Accuracy?4 of 22HOTCHIPS 2022HNPU-V2:A 46.6 FPS DNN Training Processor for Real-World Environmental Adaptation based Robust Object Detection on Mobile DevicesPromising Solution:On-device DNN Training Personalization:High Accuracy only for User-specific Task Adaptation:Performance Recover

6、y using Online TuningLocal DNN w/Opt.QuantizationPruningOriginal DNNGlobalDatasetLocalDataset1)Optimization w/Pruning/Quant.2)Low-latency DNN TuningHigh AccuracyLow AccuracyPre-trainedAdapted?TuningEdge/MobileCloud5 of 22HOTCHIPS 2022HNPU-V2:A 46.6 FPS DNN Training Processor for Real-World Environme