Prism: Mining Task-aware Domains in Non-i.i.d. IMU Data for Flexible User Perception

TL;DR

Prism addresses the challenge of non-i.i.d. IMU data in mobile user perception by automatically discovering latent, task-specific domains and training domain-aware models via an EM-based framework. It detects non-i.i.d. data, learns a shared encoder plus domain-specific classifiers on the cloud, and performs efficient on-device inference by selecting the best-fitting domain model for each test sample. Across AR and UA tasks on multiple public and large-scale SHL datasets, Prism achieves state-of-the-art FUP performance with low inference latency, outperforming semantic or clustering baselines and traditional single-model approaches. The work provides a rigorous convergence analysis, practical deployment details, and a thorough evaluation of cost and latency on real devices. Overall, Prism offers a principled, scalable solution for robust, flexible perception from non-i.i.d. IMU streams in mobile settings.

Abstract

A wide range of user perception applications leverage inertial measurement unit (IMU) data for online prediction. However, restricted by the non-i.i.d. nature of IMU data collected from mobile devices, most systems work well only in a controlled setting (e.g., for a specific user in particular postures), limiting application scenarios. To achieve uncontrolled online prediction on mobile devices, referred to as the flexible user perception (FUP) problem, is attractive but hard. In this paper, we propose a novel scheme, called Prism, which can obtain high FUP accuracy on mobile devices. The core of Prism is to discover task-aware domains embedded in IMU dataset, and to train a domain-aware model on each identified domain. To this end, we design an expectation-maximization (EM) algorithm to estimate latent domains with respect to the specific downstream perception task. Finally, the best-fit model can be automatically selected for use by comparing the test sample and all identified domains in the feature space. We implement Prism on various mobile devices and conduct extensive experiments. Results demonstrate that Prism can achieve the best FUP performance with a low latency.

Figures (10)

• Figure 1: Illustration of IMU data partition using Prism, where non-i.i.d. samples are divided into task-specific domains, denoted with different colors, rather than prior-defined subsets according to device positions or types.
• Figure 2: System architecture of Prism, where the IMU datasets are first detected whether they are non-i.i.d. and the non-i.i.d. IMU datasets is then partitioned for model training.
• Figure 3: Illustration of NID calculation, where NIs are calculated and averaged as dataset $D$ is traversed by swapping clips between $D_i^{\alpha}$ and $D_i^{\beta}$.
• Figure 4: Non-i.i.d. degree of a dataset (NID) vs perception prediction error, where a high prediction error is always with a high NID.
• Figure 5: Hyper-parameters selection results across 6 testing sets $\Gamma_i^{\text{tst}}$ for $i \in [1,6]$ on the UA task of UCI dataset.