Bidirectional Uncertainty-Driven Active Learning for Data-Efficient Permanent Magnetic Localization

Qiyu Yang

doi:10.1109/tim.2026.3742036

Bidirectional Uncertainty-Driven Active Learning for Data-Efficient Permanent Magnetic Localization

2026-10-09

0

PRE

AI

DOI：10.1109/tim.2026.3742036

Original

Original request for help

Save

Abstract

En 中文

Uniform sampling for permanent magnetic localization (PML) inefficiently captures high-order nonlinearities and spatial variations in signal-to-noise ratio, leading to excessive data acquisition. We propose an active learning framework that integrates a bidirectional uncertainty metric (BiUM) with a mixture-of-experts (MoE) architecture. BiUM, motivated by mutual information, combines uncertainty in magnetic-field generation and pose inference to rank candidate samples. The MoE architecture sequentially trains regional experts using BiUM-guided sampling, then fine-tunes all experts jointly with a gating network. Experiments on an autonomous five-axis PML platform for tongue-motion tracking show that our method reduces the sampling time by 49%, while matching the precision of state-of-the-art methods trained on full data. The code is available at https://github.com/Wen-Roc/BiUM-Active-Learning.

Keywords:

Permanent magnetic localization (PML)

active learning (AL)

mixture-of-experts (MoE)

Bidirectional Uncertainty-Driven Active Learning for Data-Efficient Permanent Magnetic Localization

Abstract

Journal

Organization

Cited Papers