The research of Hiroki Takenaka, a third-year doctoral student in our laboratory, has been accepted as an academic paper for publication in IEEE Sensors Latters. Conventional unsupervised pre-training for action recognition using sensor data involves multiple steps and is complicated. In this paper, we experimentally demonstrated that pre-training can be simplified by a simple unsupervised learning objective such as instance classification, significantly reducing the pre-training cost for action recognition.

Takenaka, K., Sakai, S., & Hasegawa, T., "IDMatchHAR: Semi-supervised Learning for Sensor-based Human Activity Recognition using Pre-training.", IEEE Sensors Letters , Vol. 9, No. 4, April 2025. paper

Abstract

In sensor-based human activity recognition (HAR), the annotation cost for sensor data is higher compared to data, such as images. One can use semisupervised learning (semi-SL) to reduce annotation costs. This method lever-ages unlabeled datasets by assigning pseudolabels. How- ever, these methods have the issue of confirmation bias, where performance degrades due to incorrect pseudolabels. Some approaches have attempted to solve this problem by performing multistage pretraining with labeled and unlabeled data, but these methods require significant computational resources. We propose a framework called IDMatchHAR, which performs semi-SL with a single-stage pretraining process on small-scale datasets. We use instance discrimination (ID) during pretraining to learn robust feature representations applied to various subsequent semi-SL tasks. We verify the effectiveness of the proposed framework using convolutional neural networks (CNNs), such as VGG and residual network (ResNet), as well as Transformers, on HASC, WISDM, and Pamap2. Our proposed framework significantly reduces the computational cost of pretraining while demonstrating performance comparable to exceeding or that of existing semi-SL methods.