[Paper Review] Controllable Pareto Multi-Task Learning

Xi Lin, Zhiyuan Yang+2 more|arXiv (Cornell University)|Oct 13, 2020

Machine Learning and Data Classification54 references22 citations

TL;DR

This paper proposes a controllable Pareto multi-task learning framework that enables real-time trade-off control among multiple tasks using a single hypernetwork-based model. By conditioning model parameters on preference vectors, the method learns the entire Pareto front end-to-end, eliminating the need to train and store multiple models, and achieves efficient, dynamic performance adjustment across diverse MTL applications with strong empirical results on benchmark datasets.

ABSTRACT

A multi-task learning (MTL) system aims at solving multiple related tasks at the same time. With a fixed model capacity, the tasks would be conflicted with each other, and the system usually has to make a trade-off among learning all of them together. For many real-world applications where the trade-off has to be made online, multiple models with different preferences over tasks have to be trained and stored. This work proposes a novel controllable Pareto multi-task learning framework, to enable the system to make real-time trade-off control among different tasks with a single model. To be specific, we formulate the MTL as a preference-conditioned multiobjective optimization problem, with a parametric mapping from preferences to the corresponding trade-off solutions. A single hypernetwork-based multi-task neural network is built to learn all tasks with different trade-off preferences among them, where the hypernetwork generates the model parameters conditioned on the preference. For inference, MTL practitioners can easily control 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.

Motivation & Objective

• To address the challenge of conflicting tasks in multi-task learning (MTL) that require dynamic trade-offs in real-world applications.
• To eliminate the need for training and storing multiple models for different trade-off preferences, which is costly and inefficient.
• To develop a unified, end-to-end trainable framework that generates diverse Pareto-optimal solutions based on user-specified preferences.
• To enable real-time inference-time control over task performance trade-offs using a single learned model.

Proposed method

• Formulate MTL as a preference-conditioned multiobjective optimization problem, where preferences map to specific trade-off solutions on the Pareto front.
• Design a hypernetwork-based MTL architecture that generates task-specific model parameters conditioned on a preference vector input.
• Use parametric mappings from preferences to trade-off solutions, enabling continuous and controllable interpolation across the Pareto front.
• Train the entire system end-to-end using a single model, with shared representations and task-specific heads generated via the hypernetwork.
• Integrate preference embeddings and chunk embeddings to represent task-specific and shared parameters efficiently.
• Optimize the model using standard deep learning training procedures (e.g., Adam, cross-entropy, L1 loss) with balanced hyperparameters across tasks.

Experimental results

Research questions

• RQ1 Can a single MTL model learn the entire Pareto front of trade-off solutions across multiple tasks?
• RQ2 Can preference-conditioned hypernetworks enable real-time, dynamic control of task performance during inference?
• RQ3 How does the proposed method compare to training multiple separate models for different trade-offs in terms of efficiency and performance?
• RQ4 Can the method scale to large-scale MTL problems with many tasks or complex architectures?

Key findings

• The proposed method successfully learns the entire Pareto front for multi-task learning problems, enabling real-time trade-off control with a single model.
• Experiments on MultiMNIST, CityScapes, NYUv2, and CIFAR-100 with 20 tasks show consistent performance across diverse MTL tasks.
• The model achieves competitive or superior performance compared to baseline MTL models, while eliminating the need for multiple model storage.
• The method scales well to large-scale MTL problems, including those with 20 tasks, demonstrating robustness and generalization.
• Preference-based control during inference allows practitioners to dynamically adjust performance trade-offs without retraining.