Interactive Pareto Navigation for Deep Multi-Task Learning

This paper introduces Preference Pareto Exploration (PPE), a novel framework designed to simplify and enhance multi-task learning, particularly when dealing with numerous objectives. Traditional approaches often aggregate losses using weighted sums, which can fail to capture decision-maker preferences or become computationally expensive with deep learning models due to the complex geometry of the Pareto front. PPE addresses these challenges by providing an interactive exploration process that aligns with the decision maker's preferences while considering the Pareto set's geometry. It employs a predictor-corrector method: predictor steps move tangentially along the manifold of Pareto-optimal solutions based on preference, and corrector steps refine the new trade-off. To ensure efficiency and robustness, PPE avoids explicit Hessian computations by utilizing a Krylov subspace method, which relies on matrix-vector products obtainable through automatic differentiation. The method's efficacy is demonstrated on both synthetic problems and deep learning applications, showcasing its ability to navigate complex trade-off spaces effectively.