Vector Quantization Improves Multiclass Calibration in ML

A new machine learning method called Divide et Calibra uses vector quantization to improve multiclass calibration by learning heterogeneous calibration maps across the latent space. The approach addresses limitations of global methods that assume homogeneous errors and local methods that suffer from information loss during dimensionality reduction. It constructs region-specific calibration maps from shared codeword-dependent factors, using vector quantization to partition the representation space and an indexed parameterization of Dirichlet concentrations for parameter sharing. Experiments on benchmark datasets show significant improvements in local calibration, especially in sparse regions. The paper is available on arXiv under ID 2605.21060.