Evolving models for incrementally learning emerging activities
Juan Ye, Elise Callus · Journal of Ambient Intelligence and Smart Environments · 2020
Ambient Assisted Living (AAL) systems are increasingly being deployed in real-world environments and for long periods of time. This significantly challenges current approaches that require substantial setup investment and cannot account for frequent, unpredictable changes in human behaviours, healt h conditions, and sensor deployments. The state-of-the-art methodology in studying human activity recognition is cultivated from short-term lab or testbed experimentation, i.e., relying on well-annotated sensor data and assuming no change in activity models. This paper propose a technique, EMILEA, to evolve an activity model over time with new types of activities. This technique novelly integrates two recent advances in continual learning: Net2Net – expanding the architecture of a model while transferring the knowledge from the previous model to the new model and Gradient Episodic Memory – controlling the update on the model parameters to maintain the performance on recognising previously learnt activities. This technique has been evaluated on two real-world, third-party, datasets and demonstrated promising results on enhancing the learning capacity to accommodate new activities that are incrementally introduced to the model while not compromising the accuracy on old activities.