Research map / Probabilistic and scientific modeling / Molecules and proteins

Divide et Calibra: Multiclass Local Calibration via Vector Quantization

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Sparse graphical and tensor models · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VQ-based compositional local multiclass calibration, where codeword-indexed sender/receiver Dirichlet calibration parameters are shared across Voronoi cells, improving local calibration (LCE/MLCE) on CIFAR-10/100 and TissueMNIST.

Problem

Multiclass post-hoc calibration is hard: global methods miss proximity bias in low-density regions, while local methods suffer from high-dimensional data sparsity and information loss from dimensionality reduction.

Approach

Quantize contiguous latent sub-vectors with a shared codebook so each sample maps to a structured index sequence (Voronoi cell); calibration parameters are indexed by the same codewords and combined via a bilinear sender/receiver Dirichlet parameterization, so rare regions reuse well-estimated codeword factors.

Key results

• On CIFAR-10, VQ reaches LCE ≈ 0.0059 ± 0.0002 vs. ≈ 0.0079 for Local Nets (≈25% drop).
• On TissueMNIST, VQ's LCE ≈ 0.0083 ± 0.0002 vs. ≈ 0.0208 for Local Nets (≈60% drop).
• VQ shows the largest LCE gains in low effective-sample-size regions, while being competitive (not uniformly best) on ECCE and NLL.

Surprising findings

• VQ does not uniformly outperform strong global baselines on ECCE or NLL.

Limitations

• Relies on a frozen representation space; success depends on whether latent space captures miscalibration-relevant locality.
• Theory is local and conditional on regularity assumptions and stable quantization assignments.
• Does not uniformly beat strong global baselines on ECCE or NLL; evaluated mainly on in-distribution classification.

Experimental setup

Datasets: CIFAR-10 CIFAR-100 TissueMNIST Weather

Benchmarks: CIFAR-10 CIFAR-100 TissueMNIST

Metrics: LCE MLCE ECCE NLL ECE ACC

Baselines: Temperature Scaling Isotonic Regression Platt Scaling Dirichlet Calibration Structured Matrix Scaling KCal Local Nets ProCal No calibration

Builds on: Dirichlet calibration KCal Local Nets ProCal

Closest papers by meaning

• Improving Multi-Class Calibration through Normalization-Aware Isotonic Techniques (ICML 2025 · similarity 0.89)
• Reassessing How to Compare and Improve the Calibration of Machine Learning Models (ICLR 2025 · similarity 0.87)
• Beyond One-Hot Labels: Semantic Mixing for Model Calibration (ICML 2025 · similarity 0.87)
• A Benchmark Study on Calibration (ICLR 2024 · similarity 0.86)
• Coverage-Based Calibration for Post-Training Quantization via Weighted Maximum Coverage over Outlier Channels (NeurIPS 2026 · similarity 0.86)
• Generalized Venn and Venn-Abers Calibration with Applications in Conformal Prediction (ICML 2025 · similarity 0.86)