Interactive Pareto navigation for deep multi-task learning

2026-06-17

AI Summary

This work addresses the challenges in multi-task learning where increasing the number of objectives leads to substantial computational overhead and difficulty in effectively balancing trade-offs, as conventional weighted-sum approaches either fail to accurately capture user preferences or incur prohibitive costs. To overcome these limitations, the authors propose the Preference Pareto Exploration (PPE) framework, which uniquely integrates interactive Pareto navigation with deep multi-task learning. By leveraging the geometric structure of the Pareto front manifold, PPE performs tangential prediction along user-preference directions followed by a correction step to efficiently generate new solutions aligned with those preferences. Built upon a Krylov subspace-based predictor-corrector mechanism, the method requires only matrix-vector products and automatic differentiation, thereby avoiding explicit Hessian computation and repeated model retraining. Experiments demonstrate that PPE achieves efficient, robust, and accurate preference-responsive multi-objective optimization across multiple benchmark tasks.

Abstract

In multi-task learning, handling an increasing number of objectives can quickly become challenging, both in terms of the computational resources and the decision maker's capacity to choose appropriate trade-offs. A widely used approach is thus to aggregate the individual losses in a single loss function by a weighted sum. This often fails to capture either the decision maker's preferences as a result of the shape of the Pareto front, or requires multiple adjustments and computations which becomes prohibitively expensive in deep learning applications. To address these issues, we introduce a novel framework, Preference Pareto Exploration (PPE), which enforces the decision maker's preferences while accounting for the geometry of the Pareto set in an interactive exploration process. PPE is based on a predictor-corrector method that performs predictor steps tangential to the manifold of Pareto-optimal solutions, following the decision maker's preference. The subsequent corrector step results in a new trade-off reflecting this preference. To avoid explicit Hessian computations when characterizing the tangent space of the manifold, we employ a Krylov subspace method that relies solely on matrix-vector products. These products can be efficiently obtained via automatic differentiation, ensuring both efficiency and robustness throughout the optimization process. The method's functionality and performance are demonstrated using both toy problems and examples from deep learning.

Problem

• multi-task learning
• Pareto front
• preference elicitation
• trade-off
• deep learning

Innovation

• Preference Pareto Exploration
• multi-task learning
• Pareto navigation
• predictor-corrector method
• Krylov subspace