Bachelor/Master Thesis: Meta-Learning for Personalized Human Activity Recognition

If you are interested in this topic, please contact Max Burzer (burzer@teco.edu)

Background

Human Activity Recognition (HAR) using data from wearable or smartphone sensors (e.g., accelerometers, gyroscopes) is an increasingly active research area with broad applications in healthcare, sports, mobile computing, and human-computer interaction.

Most HAR systems are built on generalized models trained across diverse populations. However, human actions vary greatly between individuals due to factors such as age, gender, and lifestyle. This variability often leads to reduced performance when generalized models are applied to new subjects, especially when only limited personal data is available.

To address this limitation, various personalization methods have been explored. While transfer learning and few-shot learning offer some adaptation capabilities [1][2], they remain limited in flexibility, efficiency, and overall performance. A more promising approach is meta-learning, which trains models to rapidly adapt to new tasks, in our case new subjects, with minimal data. Although prior work has applied meta-learning to personalized HAR [3], these efforts have primarily relied on MAML [4], which involves bi-level optimization and is therefore computationally inefficient.

This thesis proposes the use of Neural Processes [5], an efficient black-box meta-learning approach, for personalized HAR. Its performance will be evaluated against existing methods in terms of adaptability, classification accuracy, and computational efficiency.

Tasks

  • Review literature on personalization strategies for HAR, including transfer learning, few-shot learning, and meta-learning.
  • Design, implement, and train a meta-learning approach for personalized HAR.
  • Evaluate the approach in terms of classification performance, adaptability to new subjects, and computational efficiency.
  • Benchmark results against generalized models and transfer learning baselines.

Requirements

  • Solid understanding of deep learning concepts.
  • Proficient programming skills in Python and PyTorch.
  • Experience in working with time series data is a plus but can be acquired during the thesis.

References

  1. Bursa, S., Incel, O., & Isiklar Alptekin, G. (2023). Personalized and Motion-Based Human Activity Recognition with Transfer Learning and Compressed Deep Learning Models. Computers and Electrical Engineering, 109, 108777.
  2. Fan, F., Gu, Y., Shen, J., Dong, F., & Chen, Y. (2023). FewShotBP: Towards Personalized Ubiquitous Continuous Blood Pressure Measurement. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 7(3), 1–39.
  3. Wijekoon, A., & Wiratunga, N.. (2020). Learning-to-Learn Personalised Human Activity Recognition Models.
  4. Chelsea Finn, Pieter Abbeel, & Sergey Levine (2017). Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. CoRR, abs/1703.03400.
  5. Marta Garnelo, Jonathan Schwarz, Dan Rosenbaum, Fabio Viola, Danilo J. Rezende, S. M. Ali Eslami, & Yee Whye Teh (2018). Neural Processes. CoRR, abs/1807.01622.