Gated Adaptation for Continual Learning in Human Activity Recognition
In arXiv:2603.10046, a novel framework for continual learning that is parameter-efficient has been introduced for human activity recognition (HAR) utilizing wearable sensors. This framework tackles the issue of catastrophic forgetting in domain-incremental HAR by employing channel-wise gated modulation on frozen pretrained representations. This allows on-device models to adjust to new users while keeping sensitive information secure and off the cloud. The main idea is to focus on feature selection instead of feature generation, striking a balance between accommodating new tasks and maintaining stability with existing knowledge. This method is especially pertinent for IoT applications, including remote health monitoring, elderly care, and smart home automation.
Key facts
- arXiv:2603.10046 proposes a continual learning framework for HAR
- Framework uses channel-wise gated modulation on frozen pretrained representations
- Addresses catastrophic forgetting in domain-incremental HAR
- Enables adaptation to new subjects without cloud data transmission
- Balances plasticity and stability for on-device models
- Relevant for IoT applications like remote health monitoring and elderly care
Entities
—