Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CODA: A COst-efficient Test-time Domain Adaptation Mechanism for HAR

Minghui Qiu, Yandao Huang, Lin Chen, Lu Wang, Kaishun Wu

TL;DR

CODA tackles dynamic domain drift in HAR by presenting a cost-efficient, test-time adaptation framework that operates entirely on-device. It pairs an instance-based cache with Importance-Weighted Active Learning and a clustering loss, augmented by retentive time weights, to update predictions without trainable parameters. The approach integrates smoothly with NN-based HAR and demonstrates robust performance across phone, watch, and integrated-sensor datasets under partial feedback, while maintaining low computational overhead. This work highlights the practicality of unobtrusive, continual HAR adaptation suitable for resource-constrained mobile devices.

Abstract

In recent years, emerging research on mobile sensing has led to novel scenarios that enhance daily life for humans, but dynamic usage conditions often result in performance degradation when systems are deployed in real-world settings. Existing solutions typically employ one-off adaptation schemes based on neural networks, which struggle to ensure robustness against uncertain drifting conditions in human-centric sensing scenarios. In this paper, we propose CODA, a COst-efficient Domain Adaptation mechanism for mobile sensing that addresses real-time drifts from the data distribution perspective with active learning theory, ensuring cost-efficient adaptation directly on the device. By incorporating a clustering loss and importance-weighted active learning algorithm, CODA retains the relationship between different clusters during cost-effective instance-level updates, preserving meaningful structure within the data distribution. We also showcase its generalization by seamlessly integrating it with Neural Network-based solutions for Human Activity Recognition tasks. Through meticulous evaluations across diverse datasets, including phone-based, watch-based, and integrated sensor-based sensing tasks, we demonstrate the feasibility and potential of online adaptation with CODA. The promising results achieved by CODA, even without learnable parameters, also suggest the possibility of realizing unobtrusive adaptation through specific application designs with sufficient feedback.

CODA: A COst-efficient Test-time Domain Adaptation Mechanism for HAR

TL;DR

CODA tackles dynamic domain drift in HAR by presenting a cost-efficient, test-time adaptation framework that operates entirely on-device. It pairs an instance-based cache with Importance-Weighted Active Learning and a clustering loss, augmented by retentive time weights, to update predictions without trainable parameters. The approach integrates smoothly with NN-based HAR and demonstrates robust performance across phone, watch, and integrated-sensor datasets under partial feedback, while maintaining low computational overhead. This work highlights the practicality of unobtrusive, continual HAR adaptation suitable for resource-constrained mobile devices.

Abstract

In recent years, emerging research on mobile sensing has led to novel scenarios that enhance daily life for humans, but dynamic usage conditions often result in performance degradation when systems are deployed in real-world settings. Existing solutions typically employ one-off adaptation schemes based on neural networks, which struggle to ensure robustness against uncertain drifting conditions in human-centric sensing scenarios. In this paper, we propose CODA, a COst-efficient Domain Adaptation mechanism for mobile sensing that addresses real-time drifts from the data distribution perspective with active learning theory, ensuring cost-efficient adaptation directly on the device. By incorporating a clustering loss and importance-weighted active learning algorithm, CODA retains the relationship between different clusters during cost-effective instance-level updates, preserving meaningful structure within the data distribution. We also showcase its generalization by seamlessly integrating it with Neural Network-based solutions for Human Activity Recognition tasks. Through meticulous evaluations across diverse datasets, including phone-based, watch-based, and integrated sensor-based sensing tasks, we demonstrate the feasibility and potential of online adaptation with CODA. The promising results achieved by CODA, even without learnable parameters, also suggest the possibility of realizing unobtrusive adaptation through specific application designs with sufficient feedback.
Paper Structure (22 sections, 6 equations, 18 figures, 1 table, 2 algorithms)

This paper contains 22 sections, 6 equations, 18 figures, 1 table, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background and Motivation
  3. Domain Drifting in Human Activity
  4. Domain Adaptation for Human Activity Recognition.
  5. Active Learning for Candidates Evaluation.
  6. Online Adaptation With Active Learning
  7. Importance-Weighted Active Learning
  8. Preliminaries
  9. Skeleton of IWAL
  10. Rejection Threshold
  11. IWAL for Nearest Neighbor Model
  12. Clustering Loss Function
  13. Bounded Kernel
  14. Retentive Weights.
  15. Evaluation
  16. ...and 7 more sections

Figures (18)

  • Figure 1: Unexpected degradations of the system in the wild probably result from the uncertain drifting of the data distribution (in both durations and directions).
  • Figure 4: Adaptation pipeline in CODA (at time $T$).
  • Figure 6: Practical online domain adaptation. The underlined ratios are taken for later experiments as Minimum Feedback Ratio.
  • Figure 7: Practical Online Domain Adaptation (Augmented with MetaSense). The size of a marker only represents the relative value of standard deviation.
  • Figure 8: Comparison of Latency (ms) by prediciton and adaptation. The models of the smartwatch are marked with the release date.
  • ...and 13 more figures