Continual learning in sensor-based human activity recognition with dynamic mixture of experts

Human activity recognition (HAR) is a key enabler for many applications in healthcare, factory automation, and smart home. It detects and predicts human behaviours or daily activities via a range of wearable sensors or ambient sensors embedded in an environment. As more and more HAR applications are deployed in the real-world environments, there is a pressing need for the ability of continually and incrementally learning new activities over time without retraining the HAR model. Recently, various continual learning techniques have been applied to HAR; however, most of them commit to a large architecture, which might not suit to devices that deploy HAR models. In addition, these techniques often require to deploy the same large architecture on the devices and cannot customise the architecture for different requirements. To tackle this challenge, we present a dynamic mixture-of-experts approach, which grows an expert for each new task and allows flexible composition of experts to suit individual needs of applications. We have empirically evaluated our technique on 4 third-party, publicly available datasets and compared with 11 state-of-the-art continual learning techniques. Our results demonstrate that our technique can achieve better or comparable performance but with much less parameter spaces and training time.