17dez2025
11:00 Master's Defense room 85 of IC2
Topic on
Combinatorial U-Curve Optimization for Auxiliary Task Learning (ATL): A Task Selection Approach
Student
Giovana Kerche Bonás
Advisor / Teacher
Marcelo da Silva Reis - Co-advisor: Marcos Medeiros Raimundo
Brief summary
In many domains, the task of interest has little labeled data, which limits the performance of Single-Task Learning (STL). Auxiliary Task Learning (ATL), a variant of Multi-Task Learning (MTL), aims to transfer information from source tasks to improve the generalization of the target task. However, poorly related sources generate negative transfer and can degrade the result. Most Multi-Task Learning (MTL) methods optimize weights or gradients for a fixed set of tasks, leaving open the critical question of which tasks to select. Exhaustive subset searches can be computationally infeasible, while domain knowledge-based heuristics are not scalable. To address this gap, we formulate task selection as a two-level combinatorial optimization problem. We propose a model-agnostic framework that uses a branch-and-bound search guided by the U-curve hypothesis to efficiently explore the search space. Our strategy prunes branches where the inclusion of a new task does not reduce the target validation loss below a threshold, operating automatically and without relying on domain knowledge. The approach was evaluated on synthetic databases with controlled correlations and in real-world spam and landmine detection scenarios. The results demonstrate that our method outperforms, on average, Single Task Learning (STL) and comparative MTL approaches, with significant gains in AUC and Average Precision, as well as lower Brier Score and competitive log-loss. The analysis confirms the framework's ability to mitigate negative handoff and identify relevant task groups, establishing data-driven selection as an essential component for ATL success.
Examination Board
Headlines:
| Marcelo da Silva Reis | IC / UNICAMP |
| Fabrício Martins Lopes | DACOM/UTFPR |
| Rafael Crivellari Saliba Schouery | IC / UNICAMP |
Substitutes:
| Esther Luna Colombini | IC / UNICAMP |
| Fernando José Von Zuben | FEEC / UNICAMP |