Differentiable Calibration Losses for Multi-Class Machine Learning Models

Descrizione

Differentiable Calibration Losses for Multi-Class Machine Learning Models Sintesi in italiano (traduzione automatica): L'ISBA (Istituto di Statistica, Biostatistica e Scienze Attuariali) cerca un candidato per un dottorato di ricerca a tempo pieno, inizialmente finanziato per 15 mesi, con possibilità di estensione. La sede è a Louvain-la-Neuve, Belgio, e il supervisore è il Prof. Olivier Caelen. Il progetto di dottorato si concentra sullo sviluppo di perdite di calibrazione differenziabili per modelli di machine learning multi-classe, affrontando la questione della calibrazione dei punteggi di probabilità nei modelli di classificazione. I requisiti includono competenze in deep learning, statistica e intelligenza artificiale affidabile. È preferibile una laurea in statistica, informatica o un campo correlato. Research Institute : ISBA (Institute of Statistics, Biostatistics and Actuarial Science) which is a part of Louvain Institute of Data Analytics and Modeling in Economics and Statistics (LIDAM). Supervisor: Prof. Olivier Caelen. Contract : Full-time doctoral fellowship, initially funded for 15 months through the UCLouvain FSR Projects scheme. The fellowship is intended to support the first stage of a longer PhD trajectory. Continuation beyond the initial funded period will depend on scientific progress, evaluation, external doctoral funding applications. Start of contract : October 2026 Campus : Louvain-la-Neuve, Belgium. Scientific fields : Deep learning, statistics, trustworthy AI. Proposed PhD topic: We are seeking a motivated PhD candidate to work on a doctoral project at the intersection of machine learning, statistics, and reliable artificial intelligence. Modern classification models, especially deep neural networks, often output probability scores. These scores are commonly interpreted as confidence levels. For example, if a model assigns a probability of 70% to a class, one would expect the prediction to be correct approximately 70% of the time. A model satisfying this property is said to be well calibrated. In practice, high predictive accuracy does not guarantee good calibration. Many modern classifiers are accurate but overconfident or underconfident. This is problematic when probability scores are used to support decisions, risk assessment, medical prediction, fraud detection, ranking systems, or any setting where the reliability of the predicted probabilities matters. The objective of this PhD project is to develop differentiable calibration losses for multi-class machine learning models. The goal is to move beyond post-hoc calibration methods and design calibration criteria that can be directly incorporated into the training of neural networks. Annuncio in inglese. Fonte: Euraxess (Commissione europea).