EA-Prune: Environment Adaptive Neural Network Pruning for Low-power Energy Harvesting Devices
TimeTuesday, July 12th6pm - 7pm PDT
LocationLevel 2 Lobby
DescriptionEnergy harvesting technology that harvests energy from ambient environment has been increasingly employed to power IoT devices embedded with AI algorithms. However, once deployed, the AI models lack the ability to adapt to changing environments in real-time. In this paper, we propose a software-hardware co-design framework consisting of reinforcement learning-based architecture search and a sequential weight pruning workflow that generates three compressed models with shared-weight. EA-Prune enables environment-aware model reconfiguration during run time which affords all models within a single-model budget. A performance predictor that analyzes communication and computation resources on the given hardware and abstract the best compression configuration.