Pareto Multi-Task Learning and Its Applications

帕累托多任務學習及其應用

Student thesis: Doctoral Thesis

View graph of relations

Author(s)

Related Research Unit(s)

Detail(s)

Awarding Institution
Supervisors/Advisors
Award date21 Sep 2020

Abstract

Multi-task learning (MTL) is a powerful method for an intelligent agent to learn multiple related tasks simultaneously. By solving the related tasks together, MTL can improve each task's performance with the knowledge shared among them. It can also reduce the inference time by conducting all tasks with a single model. For many real-world problems, however, it is often impossible to find a single best solution to optimize all the tasks, since different tasks might conflict with each other. MTL practitioners have to make a trade-off among different tasks. How to efficiently combine different tasks and make a proper trade-off among them is challenging for many MTL applications.

This thesis focuses on designing efficient algorithms to solve multi-task learning problems with different trade-off preferences among the tasks. The main contributions are summarized as follows:

First, we propose a novel Pareto multi-task learning algorithm (Pareto MTL) to find a set of well-representative Pareto solutions for a multi-task learning problem. Each Pareto solution represents a different optimal trade-off among the tasks. The proposed algorithm formulates the MTL problem as a multiobjective optimization problem, and then decomposes it into a set of constrained subproblems with different trade-off preferences. By solving these subproblems in parallel, Pareto MTL can find a set of widely distributed Pareto optimal solutions with different trade-offs among the tasks. MTL practitioners can easily select their preferred solutions from these Pareto solutions, or store and use different trade-off solutions for different situations. Experimental results indicate that the proposed algorithm can generate well-representative Pareto solutions and outperform some state-of-the-art algorithms on different MTL applications.

Secondly, we generalize the idea of Pareto MTL to solve the continual learning problem. In this problem, an intelligent agent needs to incrementally learn new tasks in sequence without forgetting the previous tasks. Different tasks could conflict with each other in the online setting when the agent has a fixed learning capacity. We propose a novel preference-guided algorithm to incorporate different trade-off preferences among tasks into the continual learning process. Our proposed algorithm reformulates the continual learning problem as an online preference-guided multiobjective optimization problem. It aims to maximize the performance on the preferred task(s) while keeping good overall performance on others. Experiments on different continual learning problems validate the effectiveness of our proposed algorithm.

The above two algorithms need to build multiple models separately for different trade-off preferences among tasks, which could be costly and undesirable for many applications. Therefore, we further propose a novel controllable Pareto multi-task learning (Controllable Pareto MTL) framework, to enable an intelligent agent to make real-time trade-offs among tasks with a single model. To be specific, we formulate the MTL as a preference-conditioned multiobjective optimization problem, with a parametric mapping from the preferences to the optimal Pareto solutions. A single hypernetwork-based multi-task neural network is proposed for learning all tasks with different trade-off preferences among them. During the inference time, practitioners can easily switch the model parameters and hence the model performance based on different trade-off preferences in real time. Experiments on different applications demonstrate that the proposed model is efficient for solving various MTL problems.

Lastly, we extend the controllable Pareto MTL algorithm for solving expensive multiobjective optimization problems, where the objective evaluation could be costly and time-consuming. To reduce the number of real objective evaluations, we build a single hypernetwork-based MTL network as the surrogate model for all objective functions. It can easily incorporate tasks relation and preference information into model building, and support preference-guided batch evaluation for optimization. The proposed algorithm is also suitable for offline optimization, where no new data would be acquired during the optimization process. Experimental results on various benchmark problems demonstrate the efficiency of the proposed algorithm.

    Research areas

  • Multi-task Learning, Multiobjective Optimization