Back to AI Research

AI Research

PI-TTA: Physics-Informed Source-Free Test-Time Adap... | AI Research

Key Takeaways

  • PI-TTA: Physics-Informed Source-Free Test-Time Adaptation for Robust Human Activity Recognition on Mobile Devices Mobile devices often struggle to maintain a...
  • Source-free test-time adaptation (TTA) is appealing for mobile and wearable sensing because it enables on-device personalization from unlabeled test streams without centralizing private data.
  • setting, widely used vision-style TTA objectives can become unstable, leading to overconfident errors, representation collapse, and catastrophic forgetting.
  • PI-TTA updates the same small parameter subset as strong source-free baselines and incurs only modest overhead, making it suitable for on-device deployment.
  • It improves long-sequence accuracy by up to 9.13% and reduces physical-violation rates by 27.5%, 24.1%, and 45.4% on USCHAD, PAMAP2, and mHealth, respectively.
Paper AbstractExpand

Source-free test-time adaptation (TTA) is appealing for mobile and wearable sensing because it enables on-device personalization from unlabeled test streams without centralizing private data. However, sensor-based human activity recognition (HAR) poses challenges that are less pronounced in standard vision benchmarks: behavioral inertial streams are temporally correlated and often exhibit within-session shifts caused by sensor rotation, placement change, and sampling-rate drift. Under this streaming non-i.i.d. setting, widely used vision-style TTA objectives can become unstable, leading to overconfident errors, representation collapse, and catastrophic forgetting. We propose PI-TTA, a lightweight source-free adaptation framework that stabilizes online updates through three physics-consistent constraints: gravity consistency, short-horizon temporal continuity, and spectral stability. PI-TTA updates the same small parameter subset as strong source-free baselines and incurs only modest overhead, making it suitable for on-device deployment. Experiments on USCHAD, PAMAP2, and mHealth under long-sequence stress tests and factorized shift protocols show that PI-TTA mitigates the severe degradation observed in confidence-driven baselines and preserves stable adaptation under sustained streaming conditions. It improves long-sequence accuracy by up to 9.13% and reduces physical-violation rates by 27.5%, 24.1%, and 45.4% on USCHAD, PAMAP2, and mHealth, respectively. These results demonstrate that physics-informed adaptation can improve accuracy, stability, and deployment reliability for real-world mobile sensing systems.

PI-TTA: Physics-Informed Source-Free Test-Time Adaptation for Robust Human Activity Recognition on Mobile Devices
Mobile devices often struggle to maintain accurate human activity recognition (HAR) because sensor data changes over time due to factors like device rotation, different body placements, and sampling-rate drift. While standard adaptation methods exist, they are often designed for image processing and can become unstable when applied to continuous, correlated sensor streams. This paper introduces PI-TTA, a framework that stabilizes on-device learning by incorporating physical laws into the adaptation process, ensuring that the model remains accurate and reliable without needing to access private user data.

The Challenge of Streaming Data

In real-world mobile sensing, data is not a collection of independent, shuffled snapshots; it is a continuous, temporally correlated stream. When a model attempts to adapt to this stream using standard "vision-style" techniques—such as minimizing prediction entropy—it can fall into a "low-entropy trap." In this state, the model becomes overconfident in its incorrect predictions, leading to a collapse in performance. Because mobile devices have limited memory and energy, they cannot easily recover from these errors, making stable, real-time adaptation a significant engineering hurdle.

How PI-TTA Works

PI-TTA addresses these instabilities by anchoring the model’s learning process in physical reality rather than relying solely on statistical confidence. It introduces three physics-consistent constraints to guide the model’s updates:

  • Gravity Consistency: Since gravity provides a constant reference, the model is constrained to keep its interpretation of the gravity vector within a physically plausible range, preventing the "drifting" often seen in other methods.

  • Short-Horizon Temporal Continuity: This term ensures that the model’s features remain consistent over short periods, preventing erratic changes in recognition.

  • Spectral Stability: This constraint helps the model remain robust even when sensor sampling rates drift, preserving the integrity of the frequency patterns associated with human movement.
    By combining these physical signals with a lightweight update mechanism, PI-TTA allows the model to personalize itself on-device without requiring access to the original training data.

Results and Performance

The researchers tested PI-TTA on three standard datasets (USCHAD, PAMAP2, and mHealth) under stressful, long-sequence conditions. The results demonstrate that PI-TTA significantly outperforms standard confidence-driven baselines. Specifically, the framework improved long-sequence accuracy by up to 9.13% and drastically reduced "physical-violation rates"—instances where the model’s internal representation of motion became physically impossible—by as much as 45.4%.

Deployment Reliability

A key focus of the research is ensuring that these improvements do not come at the cost of device performance. PI-TTA is designed to be lightweight, updating only a small subset of parameters. The study confirms that the framework is compatible with strict mobile runtime constraints, such as a 50 Hz/20 ms update deadline, making it a practical solution for real-world wearable systems that require both high accuracy and sustained, reliable operation.

Comments (0)

No comments yet

Be the first to share your thoughts!