Online Continual Learning for Embedded Devices

Real-time on-device continual learning is needed for new applications such as home robots, user personalization on smartphones, and augmented/virtual reality headsets. However, this setting poses unique challenges: embedded devices have limited memory and compute capacity and conventional machine learning models suffer from catastrophic forgetting when updated on non-stationary data streams. While several online continual learning models have been developed, their effectiveness for embedded applications has not been rigorously studied. In this paper, we first identify criteria that online continual learners must meet to effectively perform real-time, on-device learning. We then study the efficacy of several online continual learning methods when used with mobile neural networks. Continual machine learning systems have the ability to learn from ever-growing data streams (Parisi et al., 2019). In contrast, conventional machine learning algorithms typically assume that there is a static training and evaluation dataset. Continual learning has emerged as a popular research area. One of the most critical applications for continual learning is using it on embedded devices such as mobile phones, virtual/augmented reality (VR/AR) headsets, robots, vehicles, and smart appliances. VR headsets use continual learning to localize the position of the wearer within the boundary that the user has established so that the user does not collide with obstacles (O'Hagan & Williamson, 2020). AR headsets require continual learning to identify relevant objects and regions in the field of view to appropriately position virtual perceptual information. Household robotic devices need to learn the identity of the individuals, pets, and objects in the house. Typically, inference for these applications must be done within embedded devices to minimize latency, but continual on-device learning is critical to preserving privacy and security of the user. Conventional machine learning systems trained with empirical risk minimization assume that the data is independent and identically distributed (iid), which is typically enforced by shuffling the data. In continual learning, this assumption is violated, which results in catastrophic forgetting (French, 1999; Parisi et al., 2019). Hence, the continual learning research community has focused on solving this catastrophic forgetting problem in a variety of scenarios. However, most of these scenarios do not match the conditions an agent would face for embedded applications.