New Framework Unifies Edge AI Learning Scenarios

The proliferation of smart edge devices is driving demand for personalized applications, such as custom keyword spotting or adaptive health monitoring. However, most current edge devices rely on fixed inference algorithms, limiting their ability to learn and adapt locally. Existing on-device learning solutions typically support only one specific scenario, like few-shot learning, necessitating cloud retraining or specialized hardware for other tasks, which introduces latency, energy overheads, and privacy concerns. This research presents Embedder-Centric Learning (ECL), a novel framework that integrates four distinct online learning paradigms: few-shot learning for immediate customization, continual learning for knowledge accumulation, zero-shot learning for semantic data leverage, and in-context learning for broader adaptation. ECL has been successfully deployed on resource-constrained hardware, achieving new state-of-the-art results in few-shot character recognition and establishing the first hardware baselines for continual learning in keyword spotting, zero-shot spoken sentence classification, and in-context learning, all while operating within extremely low power budgets.