Active forgetting with selective labeling for multi-task learning

초록

Constructing a dataset is an indispensable process for learning. In multi-task learning, it becomes increasingly complex and costly, often leading to inaccurately assigned labels and a decline in performance. However, existing works often overlook the value of generated annotations, continuing to expand a dataset with arbitrary or less significant samples. For efficient data allocation, discarding less useful data obtained previously and prioritizing new informative data is desirable. Additionally, in multi-task learning, it is crucial to regard data task-wise since some samples might prove less beneficial for specific tasks. To address this, we propose a novel active forgetting framework for multi-task learning through selective labeling. It aims to construct a dataset composed of the most valuable data by identifying and removing uninformative task samples. To estimate the influence of the data and forget less valuable data, we introduce a multi-task unlearning strategy that leverages inter-task and intra-task dependencies to prioritize highly associated data. Extensive experimental results under various data budgets demonstrate the outstanding performance of our approach compared to existing active learning methods, showcasing the effectiveness in reducing annotation costs.

키워드