Egilea: Mario Martínez García

Izenburua: Advances in learning using privileged information for supervised classification

Zuzendariak: Jose Antonio Lozano Alonso / Iñaki Inza Cano

Eguna: 2025ko uztailaren 2an
Ordua: 10:30h
Lekua: Ada Lovelace aretoa (Informatikako fakultatea)

Abstract:

"What if machine learning models could train with extra “privileged knowledge”?  The world is increasingly data-dependent with more automatic jobs thanks to advances in the development of intelligent machines capable of learning complex patterns accurately.  However, in some scenarios, not all available information is leveraged in the machine learning process. In a clinical setting, a patient’s health record is available, but typically not all of the information is used during the learning process. For instance, some features may be discarded from the training process due to their unavailability at deployment time. These features are called privileged features and are leveraged through the Learning Using Privileged Information (LUPI) paradigm, a learning scenario that exploits privileged features as additional information for training models. The thesis collects new methodologies that incorporate the LUPI paradigm with supervised classification algorithms.  Firstly, two logistic regression-based methods learned using privileged information are presented. Secondly, a privileged knowledge distillation approach is proposed: a teacher model (trained with both regular and privileged features) transfers knowledge to a student model using only regular features. However, the teacher model may not be entirely reliable or error-free. Therefore, the proposed distillation framework includes a mechanism to guide the student correctly. Finally, a multi-task privileged framework is introduced. Thereby, one task predicts privileged features from regular ones, and another uses regular and the predicted privileged features to perform the final prediction. Moreover, this framework is also addressed using knowledge distillation techniques. It is important to note that privileged information does not inherently guarantee improved model performance. Consequently, each chapter introduces different approaches designed to maximize the advantages of privileged information and to provide a clearer understanding of its impact on model performance. All methods are validated on various datasets, demonstrating significant improvements over current state-of-the-art techniques. The work contributes both theoretical insights and practical solutions for leveraging privileged information in real-world scenarios."