The work “Divide et Calibra: Multiclass Local Calibration via Vector Quantization” with Cesare Barbera, Lorenzo Perini, Giovanni De Toni and Andrea Passerini has been accepted at NeurIPS 2026!
This paper studies how to calibrate multiclass models when calibration errors vary across the representation space. We propose a compositional approach based on vector quantization, which partitions this space and learns local calibration maps by sharing parameters across regions. This allows us to capture different calibration needs while generalizing even to regions with few data points. We finally showcase the effectiveness of our approach on benchmark datasets, improving local calibration while maintaining competitive global calibration and predictive performance.
A final thanks to all the amazing co-authors involved in this project.
See you in Paris 🗼🥐🇫🇷!