2019/04/04 by Shahriar Iqbal, Iqbal, Md Shahriar, Lars Kotthoff +4 · 1 citation
Neuroscience · Computer Science · #EEG and Brain-Computer Interfaces #Advanced Neural Network Applications #Software System Performance and Reliability
paper · pdf · doi:10.48550/arxiv.1904.02838
Modern deep neural network (DNN) systems are highly configurable with large a\nnumber of options that significantly affect their non-functional behavior, for\nexample inference time and energy consumption. Performance models allow to\nunderstand and predict the effects of such configuration options on system\nbehavior, but are costly to build because of large configuration spaces.\nPerformance models from one environment cannot be transferred directly to\nanother; usually models are rebuilt from scratch for different environments,\nfor example different hardware. Recently, transfer learning methods have been\napplied to reuse knowledge from performance models trained in one environment\nin another. In this paper, we perform an empirical study to understand the\neffectiveness of different transfer learning strategies for building\nperformance models of DNN systems. Our results show that transferring\ninformation on the most influential configuration options and their\ninteractions is an effective way of reducing the cost to build performance\nmodels in new environments.\n