E-QUARTIC：用于资源优化学习的卷积神经网络节能边缘集成.pdf

1、System Energy Efficiency Labseelab.ucsd.eduLe Zhang,Onat Gungor,Flavio Ponzina,Tajana RosingContact:Flavio PonzinaPostdoctoral scholarfponzinaucsd.eduThe unstoppable IoT market2Source:IoT Analytics linkEmpowering the edge3EdgeCloudComputing at the edgeDATADATAApplicationsHealthcareVirtual RealitySpa

2、ce&DefenseIndustrial IoTWhy edge computing?Autonomous SystemsSecurity and PrivacyPerformance and EfficiencyEdge Artificial Intelligence(AI)-Small memories-Limited computing resources-Small batteriesSoftwareHardware-Memory-intensity-Compute-intensity-Energy-intensityResearch efforts-Software Optimiza

3、tion-Hardware Optimization-SW/HW co-designEdge AI challenges4High accuracy with limited resources!Deploy Artificial Intelligence in mobile devices5Energy Harvesting(EH)IoTVery limited energy budgetsCan we empower EH IoT with AI capabilities?Dynamic energy availabilityConstrained HW resourcesHarveste

4、rs extract solar,piezoelectric,or thermal energyThey only get limited energy from the environmentLimited memory,storage(KB-MB),and compute resourcesSmall batteries can provide more stable energy sourcesExtracted energy profiles have a dynamic shape 6Optimizing convolutional neural networks for ultra

5、-low power embedded systemsQuantization and Pruning7Floating-point valuesFixed-point valuesReducing operands bitwidth(e.g.,from fp32 to int8)Integer arithmetic improves energy efficiencyIf too aggressive,may result in accuracy degradationQuantizationQuantization ACM TIST23Removes model parameters(e.

6、g.,weights)Fine-grain removes individual weightsCoarse-grain removes entire filtersPruning IEEE TPAMI24Fine-grain pruningCoarse-grain pruningCheng et al.,“A Survey on Deep Neural Network Pruning:Taxonomy,Comparison,Analysis,and Recommendations”,IEEE TPAMI 2024Rokh et al.,“A Comprehensive Survey on M