Multi-task Learning Combined with RL-based Weight Search and MC Dropout
Sicong Liang, Jing Luo, Rongquan Yang · 2021 3rd International Conference on Machine Learning, Big Data and Business Intelligence (MLBDBI) · 2021
Multi-Task Learning (MTL) is a very encouraging research direction in machine learning, which has been applied in many fields successfully. However, designing a good MTL loss function and verifying the robustness of MTL model are two key challenges in this field. In this paper, we adopt RL-based weight search method to find a set of good weights in MTL loss function to help to improve the training optimization process. To improve the robustness of MTL model, especially for out-of-distribution (OoD) detection, epistemic uncertainty of MTL model is estimated via Monte Carlo dropout (MC Dropout) in this paper. Empirical studies show the proposed methods combined with MTL model is effective.