AI-POWERED INSIGHTS

Rigorous Guarantees for Stability

Our theoretical analysis demonstrates why channel-wise gating yields improved stability. By restricting adaptation to bounded, diagonal modulation of frozen backbone features, updates induced by new subjects result in limited functional drift on previously seen data. Theorem 1 establishes 'Bounded Feature Drift,' showing that gating limits representational changes to a multiplicative factor. Theorem 2 further decomposes 'Bounded Logit Drift,' separating gate-induced and classifier-induced changes, ensuring that even significant classifier adaptation does not severely impact prior knowledge. Corollary 1 provides a 'Margin-Based Forgetting Guarantee,' indicating that predictions are preserved if logit drift is less than half the classification margin. This structured adaptation prevents uncontrolled shifts in the feature space, a key differentiator from unconstrained linear transformations.

Our empirical evaluations across PAMAP2, UCI-HAR, and DSA datasets consistently demonstrate that gated adaptation achieves superior stability-plasticity tradeoffs. On PAMAP2, our method reduced forgetting to 16.2% (from 39.7% with a trainable backbone) while boosting final accuracy to 77.7% (from 56.7%). This significantly outperforms regularization-based (EWC, LwF) and architectural (HAT) baselines without requiring data replay or task-specific regularization. We found that freezing the backbone itself (Pretrained, no gates) dramatically reduces forgetting (17.5% FM on PAMAP2) but also reduces learning accuracy. The addition of gates to the frozen backbone then restores plasticity while maintaining stability, proving their critical role.

Practical Advantages for IoT Deployments

The proposed gated adaptation framework offers substantial advantages for resource-constrained IoT edge devices. By updating less than 2% of total model parameters per task, it dramatically reduces computational cost and energy consumption compared to full fine-tuning. The use of a frozen backbone ensures a stable representation, eliminating the need for transmitting sensitive raw sensor data to the cloud for retraining, which is crucial for privacy. Our replay-free approach avoids the memory overhead and privacy concerns associated with storing past samples, making it ideal for devices with limited storage and strict privacy regulations. This approach enables on-device continual personalization, allowing HAR models to adapt to individual users' movement characteristics without compromising the overall system integrity or user data.