Exploring Large Language Models in Active Learning for Annotating Physical Sensing Data

Active learning minimizes annotation costs by lever-aging a model trained using a small bootstrap amount of labeled data to quantify the uncertainty of predictions and then delegate those samples for annotation to a human oracle. This work explores the idea of replacing the human-in-the-loop with Large Language Models (LLMs) for physical sensing data. In this paper, we systematically study the uncertainty of models and the corresponding accuracy of responses from the LLM. To perform the study in a principled manner, we used two state-of-the-art LLMs, GPT-4 and Llama-3.2, without fine-tuning or sophisticated prompt engineering. Experiments on benchmark HAR datasets, MotionSense and HHAR, demonstrate the overall idea’s viability and highlight the key challenges.