Yingya Li, Timothy Miller, Steven Bethard, Guergana Savova
{"title":"Identifying task groupings for multi-task learning using pointwise V-usable information.","authors":"Yingya Li, Timothy Miller, Steven Bethard, Guergana Savova","doi":"10.1016/j.jbi.2025.104881","DOIUrl":null,"url":null,"abstract":"<p><strong>Objective: </strong>Even in the era of Large Language Models (LLMs) which are claimed to be solutions for many tasks, fine-tuning language models remains a core methodology used in deployment for a variety of reasons - computational efficiency and performance maximization among them. Fine-tuning could be single-task or multi-task joint learning where the tasks support each other thus boosting their performance. The success of multi-task learning can depend heavily on which tasks are grouped together. Naively grouping all tasks or a random set of tasks can result in negative transfer, with the multi-task models performing worse than single-task models. Though many efforts have been made to identify task groupings and to measure the relatedness among different tasks, it remains a challenging research topic to define a metric to identify the best task grouping out of a pool of many potential task combinations. We propose such a metric.</p><p><strong>Methods: </strong>We propose a metric of task relatedness based on task difficulty measured by pointwise V-usable information (PVI). PVI is a recently proposed metric to estimate how much usable information a dataset contains given a model. We hypothesize that tasks with not statistically different PVI estimates are similar enough to benefit from the joint learning process. We conduct comprehensive experiments to evaluate the feasibility of this metric for task grouping on 15 NLP datasets in the general, biomedical, and clinical domains. We compare the results of the joint learners against single learners, existing baseline methods, and recent large language models, including Llama and GPT-4.</p><p><strong>Results: </strong>The results show that by grouping tasks with similar PVI estimates, the joint learners yielded competitive results with fewer total parameters, with consistent performance across domains.</p><p><strong>Conclusion: </strong>For domain-specific tasks, finetuned models may remain a preferable option, and the PVI-based method of grouping tasks for multi-task learning could be particularly beneficial. This metric could be wrapped in the overall recipe of fine-tuning language models.</p>","PeriodicalId":15263,"journal":{"name":"Journal of Biomedical Informatics","volume":" ","pages":"104881"},"PeriodicalIF":4.5000,"publicationDate":"2025-09-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Journal of Biomedical Informatics","FirstCategoryId":"3","ListUrlMain":"https://doi.org/10.1016/j.jbi.2025.104881","RegionNum":2,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"2025/7/16 0:00:00","PubModel":"Epub","JCR":"Q2","JCRName":"COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS","Score":null,"Total":0}
引用次数: 0
Abstract
Objective: Even in the era of Large Language Models (LLMs) which are claimed to be solutions for many tasks, fine-tuning language models remains a core methodology used in deployment for a variety of reasons - computational efficiency and performance maximization among them. Fine-tuning could be single-task or multi-task joint learning where the tasks support each other thus boosting their performance. The success of multi-task learning can depend heavily on which tasks are grouped together. Naively grouping all tasks or a random set of tasks can result in negative transfer, with the multi-task models performing worse than single-task models. Though many efforts have been made to identify task groupings and to measure the relatedness among different tasks, it remains a challenging research topic to define a metric to identify the best task grouping out of a pool of many potential task combinations. We propose such a metric.
Methods: We propose a metric of task relatedness based on task difficulty measured by pointwise V-usable information (PVI). PVI is a recently proposed metric to estimate how much usable information a dataset contains given a model. We hypothesize that tasks with not statistically different PVI estimates are similar enough to benefit from the joint learning process. We conduct comprehensive experiments to evaluate the feasibility of this metric for task grouping on 15 NLP datasets in the general, biomedical, and clinical domains. We compare the results of the joint learners against single learners, existing baseline methods, and recent large language models, including Llama and GPT-4.
Results: The results show that by grouping tasks with similar PVI estimates, the joint learners yielded competitive results with fewer total parameters, with consistent performance across domains.
Conclusion: For domain-specific tasks, finetuned models may remain a preferable option, and the PVI-based method of grouping tasks for multi-task learning could be particularly beneficial. This metric could be wrapped in the overall recipe of fine-tuning language models.
期刊介绍:
The Journal of Biomedical Informatics reflects a commitment to high-quality original research papers, reviews, and commentaries in the area of biomedical informatics methodology. Although we publish articles motivated by applications in the biomedical sciences (for example, clinical medicine, health care, population health, and translational bioinformatics), the journal emphasizes reports of new methodologies and techniques that have general applicability and that form the basis for the evolving science of biomedical informatics. Articles on medical devices; evaluations of implemented systems (including clinical trials of information technologies); or papers that provide insight into a biological process, a specific disease, or treatment options would generally be more suitable for publication in other venues. Papers on applications of signal processing and image analysis are often more suitable for biomedical engineering journals or other informatics journals, although we do publish papers that emphasize the information management and knowledge representation/modeling issues that arise in the storage and use of biological signals and images. System descriptions are welcome if they illustrate and substantiate the underlying methodology that is the principal focus of the report and an effort is made to address the generalizability and/or range of application of that methodology. Note also that, given the international nature of JBI, papers that deal with specific languages other than English, or with country-specific health systems or approaches, are acceptable for JBI only if they offer generalizable lessons that are relevant to the broad JBI readership, regardless of their country, language, culture, or health system.