arXiv cs.AI· Cesare Barbera, Lorenzo Perini, Giovanni De Toni, Andrea Passerini, Andrea Pugnana·· 14 小时前AI 评分18
Divide et Calibra：基于向量量化的多类局部校准方法
Divide et Calibra: Multiclass Local Calibration via Vector Quantization
AI 导读
研究者提出 Divide et Calibra，一种基于向量量化（VQ）的多类校准方法，通过对表示空间做结构化划分并共享 Dirichlet 浓度参数，学习区域特定的异构校准映射。实验显示该方法在基准数据集上显著提升局部校准效果，同时保持有竞争力的全局校准与预测性能。论文已被 NeurIPS 2026 接收。
正文 · 原文
Abstract:Accurate and well-calibrated Machine Learning (ML) models are mandatory in high-stakes settings, yet effective multiclass calibration remains challenging: global approaches assume calibration errors are homogeneous across the latent space, while local methods often rely on representation-space dimensionality reduction, discarding latent directions relevant to miscalibration. To address these issues, we propose a compositional approach to multiclass calibration, where region-specific calibration maps are constructed from shared factors. We instantiate this idea via Vector Quantization (VQ), which induces a structured partition of the representation space, and an indexed parameterization of Dirichlet concentrations that enables parameter sharing across regions. Our approach learns heterogeneous calibration maps that generalize well even to sparse regions of the latent space. Experiments on benchmark datasets show significant improvements in local calibration while maintaining competitive global calibration and predictive performance.
| Comments: | Accepted at the Fortieth Annual Conference on Neural Information Processing System (NeurIPS 2026) |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Machine Learning (stat.ML) |
| MSC classes: | 68T05 |
| Cite as: | arXiv:2605.21060 [cs.LG] |
| (or arXiv:2605.21060v2 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2605.21060 arXiv-issued DOI via DataCite |
Submission history
From: Cesare Barbera [view email]
[v1]
Wed, 20 May 2026 11:44:55 UTC (3,612 KB)
[v2]
Mon, 26 Oct 2026 17:35:25 UTC (3,768 KB)
来源：arXiv cs.AI · arxiv.org