24set2025
10:00 Master's Defense IC3 Auditorium
Topic on
An Algorithm for Multi-Objective Optimization in Deep Learning Using Weighted Sum Adaptation
Student
Aline Cavalca Carvalho Soares de Azevedo
Advisor / Teacher
Marcos Medeiros Raimundo
Brief summary
The presence of multiple conflicting objectives is a significant optimization challenge in machine learning and is commonly encountered in various problems, such as those related to fairness, imbalanced classes, reinforcement learning, multi-task learning, or multi-class classification. Using a single loss function is often insufficient to optimize these objectives simultaneously due to the inherent trade-offs between them. To address this, Multi-Objective Optimization (MOO) methods offer an effective alternative by allowing the learning process to explore and balance conflicting goals. Therefore, this research proposes the development of a new MOO algorithm called the Multi-Objective Learning Algorithm (MOLA), based on a weighted sum approach. MOLA is model-agnostic and was designed to work with both machine learning and deep learning, resulting in a set of models representing different trade-offs between objectives, from which users can select the most appropriate one for a given application or combine them using ensemble strategies. Additionally, the study implements a comprehensive framework, MachineMOO, that integrates MOO methods into machine learning workflows. To facilitate its adoption, it is made available as a Python package compatible with popular machine learning libraries (e.g., Scikit-learn, TensorFlow, and PyTorch). And while the framework is applicable to general machine learning problems, it is validated in this work through applications related to social justice and equity.
Examination Board
Headlines:
| Marcos Medeiros Raimundo | IC / UNICAMP |
| Petra Maria Bartmeyer | IMECC / UNICAMP |
| Leonardo Tomazeli Duarte | FCA/UNICAMP |
Substitutes:
| Christiano Lyra Filho | FEEC / UNICAMP |
| Anderson de Rezende Rocha | IC / UNICAMP |