DiBS-MTL: Transformation-Invariant Multitask Learning with Direction Oracles
This addresses a key challenge in multitask learning for AI practitioners, offering a robust solution to task scaling issues, though it builds incrementally on prior bargaining-inspired methods.
The paper tackles the problem of task domination in multitask learning due to arbitrary scaling of task losses, by proposing DiBS-MTL, a method based on a bargaining theory solution that is invariant to such transformations. It proves convergence for nonconvex settings and shows competitive performance on benchmarks, with robustness to transformations that degrade existing methods.
Multitask learning (MTL) algorithms typically rely on schemes that combine different task losses or their gradients through weighted averaging. These methods aim to find Pareto stationary points by using heuristics that require access to task loss values, gradients, or both. In doing so, a central challenge arises because task losses can be arbitrarily, nonaffinely scaled relative to one another, causing certain tasks to dominate training and degrade overall performance. A recent advance in cooperative bargaining theory, the Direction-based Bargaining Solution (DiBS), yields Pareto stationary solutions immune to task domination because of its invariance to monotonic nonaffine task loss transformations. However, the convergence behavior of DiBS in nonconvex MTL settings is currently not understood. To this end, we prove that under standard assumptions, a subsequence of DiBS iterates converges to a Pareto stationary point when task losses are possibly nonconvex, and propose DiBS-MTL, a computationally efficient adaptation of DiBS to the MTL setting. Finally, we validate DiBS-MTL empirically on standard MTL benchmarks, showing that it achieves competitive performance with state-of-the-art methods while maintaining robustness to nonaffine monotonic transformations that significantly degrade the performance of existing approaches, including prior bargaining-inspired MTL methods. Code available at https://github.com/suryakmurthy/dibs-mtl.