[공지사항]을 빙자한 안부와 근황 
Show more

Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Identifying Task Groupings for Multi-Task Learning Using Pointwise V-Usable Information

Created by
  • Haebom

Author

Yingya Li, Timothy Miller, Steven Bethard, Guergana Savova

Outline

This paper emphasizes the importance of task grouping in multi-task learning and proposes a novel metric to identify optimal task groups by measuring the relatedness between tasks. In particular, we present a method to measure the difficulty of tasks based on pointwise V-usable information (PVI) and group tasks with statistically similar PVI estimates. Through experiments on 15 NLP datasets from general, biomedical, and clinical domains, we verify the effectiveness of the proposed method compared to existing methods including Llama 2 and GPT-4. The experimental results show that tasks grouped based on PVI achieve competitive performance with fewer parameters and show consistent performance across domains.

Takeaways, Limitations

•
Takeaways:
◦
We experimentally demonstrate that PVI-based task grouping is effective in improving the performance of multi-task learning.
◦
Achieve competitive performance with existing methods and large-scale language models even with a small number of parameters.
◦
Consistent performance across a variety of domains (general, biomedical, clinical).
◦
Introducing new metrics for measuring work relevance.
•
Limitations:
◦
Further research is needed on the computational cost and efficiency of PVI calculations.
◦
Further validation of the generalization performance of the proposed indicator is needed.
◦
There is a possibility of bias towards certain domains or types of work.
◦
Additional experiments on different task types and datasets are needed.
👍
Made with Slashpage