HMC: A Hybrid Reinforcement Learning Based Model Compression for Healthcare Applications

摘要

Artificial intelligence (AI) healthcare applications to optimize workflows, reduce costs while focusing on patient care are on the rise. While deeper and wider neural networks are designed for complex healthcare applications, model compression is poised to be an effective way to deploy networks on medical devices that often have hardware and speed constraints. Most state-of-the-art model compression techniques require a resource centric manual process that explores a large space to find a trade-off solution between model size and accuracy. Recently, reinforcement learning (RL) approaches are proposed to automate such hand-crafted process. However, most RL model compression algorithms are model-free, meaning a very long time to train such RL agents due to the huge state space. In this work, we develop a hybridRL model compression (HMC) method that integrates model-based and model-free RL approaches. We demonstrate our method on a wide range of imaging data on healthcare related model architectures. Compared to model-free RL approaches, our results show that HMC method reduces the training time significantly, exhibits better generalization capabilities across different data sets, and preserves comparable model compression performance.