A Light-Weight Deployment Methodology for DNN Re-training in Resource-Constraint Scenarios.

Wireless smart devices with restricted resources in the IoT may need to update DNNs due to environmental changes. In order to alleviate the computational complexity of DNN updates and stringent requirements on the size and distribution of datasets when re-training, Deep Transfer Learning (DTL) is proposed for transferring the knowledge that DNNs can learn from large standard datasets and reducing the number of DNN layers that need to be re-trained. However, previous work has rarely reconciled the needs of the computational process with the advantages of the computational platform, resulting in sub-optimal performance of the system. To address this problem, we propose a Light-weight Deployment Methodology, which targets for agile deployment of DNN re-training in resource-constraint scenarios. We design a Hybrid Precision Light-weight Strategy to distinguish the general feature extractor layer from the others so that the different light-weight mechanisms are utilized for efficient computing. Besides analyzing the system performance from aspects of information loss, computational throughput, and resource utilization, our design is able to generate a system configuration guidance with NSGA-II, which compromises DNN performance and system efficiency. The evaluation shows the throughput of the computation module guided by NSGA-II reach 39.30 and 35.04 GOPs respectively for each quantization mechanism, with the relative errors of only 2.52% and 4.60% from the theoretical values. And the model size is reduced by 34.72% without accuracy loss.