AutoSeM: Automatic Task Selection and Mixing in Multi-Task Learning

Proceedings of NAACL-HLT 2019, pages 3520–3531 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3520 AUTOSEM: Automatic Task Selection and Mixing in Multi-Task Learning Han Guo Ramakanth Pasunuru Mohit Bansal UNC Chapel Hill {hanguo, ram, mbansal}@cs.unc.edu Abstract Multi-task learning (MTL) has achieved suc- cess over a wide range of problems, where the goal is to improve the performance of a primary task using a set of relevant auxiliary tasks. However, when the usefulness of the auxiliary tasks w.r.t. the primary task is not known a priori, the success of MTL models de- pends on the correct choice of these auxiliary tasks and also a balanced mixing ratio of these tasks during alternate training. These two problems could be resolved via manual intu- ition or hyper-parameter tuning over all com- binatorial task choices, but this introduces in- ductive bias or is not scalable when the number of candidate auxiliary tasks is very large. To address these issues, we present AUTOSEM, a two-stage MTL pipeline, where the first stage automatically selects the most useful auxiliary tasks via a Beta-Bernoulli multi-armed ban- dit with Thompson Sampling, and the sec- ond stage learns the training mixing ratio of these selected auxiliary tasks via a Gaussian Process based Bayesian optimization frame- work. We conduct several MTL experiments on the GLUE language understanding tasks, and show that our AUTOSEM framework can successfully find relevant auxiliary tasks and automatically learn their mixing ratio, achiev- ing significant performance boosts on several primary tasks. Finally, we present ablations for each stage of AUTOSEM and analyze the learned auxiliary task choices. 1 Introduction Multi-task Learning (MTL) (Caruana, 1997) is an inductive transfer mechanism which lever- ages information from related tasks to improve the primary model’s generalization performance. It achieves this goal by training multiple tasks in parallel while sharing representations, where the training signals from the auxiliary tasks can help improve the performance of the primary task. Multi-task learning has been applied to a wide range of natural language processing prob- lems (Luong et al., 2015; Pasunuru and Bansal, 2017; Hashimoto et al., 2017; Ruder et al., 2017b; Kaiser et al., 2017; McCann et al., 2018). Despite its impressive performance, the design of a multi- task learning system is non-trivial. In the con- text of improving the primary task’s performance using knowledge from other auxiliary tasks (Lu- ong et al., 2015; Pasunuru and Bansal, 2017), two major challenges include selecting the most rel- evant auxiliary tasks and also learning the bal- anced mixing ratio for synergized training of these tasks. One can achieve this via manual intuition or hyper-parameter tuning over all combinatorial task choices, but this introduces human inductive bias or is not scalable when the number of candi- date auxiliary tasks is considerable. To this end, we present AUTOSEM, a two-stage Bayesian op- timization pipeline to this problem. In our AUTOSEM framework1, the first stage addresses automatic task selection from a pool of auxiliary tasks. For this, we use a non-stationary multi-armed bandit …

Content truncated to 500 words