SelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data
summary
Combines teacher-student self-training with multi-task self-supervision: train a teacher on labeled data, generate pseudo-labels on large unlabeled sensor streams, filter high-confidence samples, pre-train a student on augmented transformation tasks and HAR labels, then fine-tune on labeled data. Results show up to 12% F1 improvement and outperform prior semi-supervised baselines on 5 of 6 datasets; achieves similar performance on HHAR with ~10× less labeled data and no extra inference cost.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 0:18What is HR
- 0:34Challenges
- 1:06Objectives
- 1:21Web Method
- 1:33Main Idea
- 2:40Results
- 4:01Results in Limited Labor Data
- 4:51Conclusion
- 5:31Outro
related talks
06mA Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition
37mActive Teaching, Human Learning
06mWeakly Supervised Multi-Task Representation Learning for Human Activity Analysis Using Wearables
30mActive Learning and Human-in-the-Loop Machine Learning
53mSelf-Supervised Learning in Transformer Models for Birdsong
52mPeer-Predictive Self-Training for Language Model Reasoning
51mHandling the problem of hand-labeled training data with data programming...
06mCross-Dataset Activity Recognition via Adaptive Spatial-Temporal Transfer Learning