University of Twente Student Theses
The Effective Use of Limited Feedback to Personalize a Mixed Model for Human Activity Recognition
Huveneers, Maartje (2024) The Effective Use of Limited Feedback to Personalize a Mixed Model for Human Activity Recognition.
|
PDF
5MB |
| Abstract: | HAR (Human Activity Recognition) is an upcoming field of technology that can potentially decrease the workload of healthcare workers by recognizing abnormal behaviour. However, existing HAR systems often lack personalization, which can lead to reduced performance. This thesis proposes a new framework to personalize a mixed model, consisting of a general and personal model, using dynamic classifier selection and novelty detection. The personal model adapts to the subject over time by receiving feedback from annotators. With the goal of reducing annotators’ efforts, LF (limited feedback) was applied, which allows annotators to select from a limited list of classes rather than the entire set, and Active Learning. The experiments were conducted on three toy problem datasets and one HAR dataset. The results show that the framework performs similarly or better than the personal model in both datasets. When the personal model lacks performance, the framework can improve its performance, although it is occasionally outperformed by the general model. Additionally, the comparison of five LF techniques (Correct Hard, Sampled Hard, Modified Soft, Unbiased Risk Estimator and Deep Naive Partial Label learning) showed that all techniques, except for Correct Hard, performed similarly to a model trained with full feedback. This indicates that the number of classes presented to the annotator could be reduced while yielding the same learning gain. However, all models suffered from catastrophic forgetting, which led to an overall low performance and could have reduced the difference in results. Next to that, entropy-based sampling was compared to random sampling but showed no improvement in reducing the number of samples required for optimal performance for both the toy problem and the HAR dataset. While the methods were tested on two diverse datasets, future work should experiment with different datasets, subjects and models to see if the results generalize. Additionally, further research should explore different dynamic classifier selection and novelty detection methods to improve the framework. Lastly, other Active Learning strategies should be tested to determine whether they can reduce the number of samples required for optimal performance. |
| Item Type: | Essay (Master) |
| Clients: | Nedap, Groenlo, The Netherlands |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 54 computer science |
| Programme: | Interaction Technology MSc (60030) |
| Link to this item: | https://purl.utwente.nl/essays/98551 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page