Reproduce “Learning Kernels by Alignment for Multiclass Bayes Classification” on OpenResearch

About this paper

Hollan Haule, Javier Escudero

Learning Kernels by Alignment for Multiclass Bayes Classification recasts Collaborative Learning and Inference (CLaI) as a supervised kernel-alignment method with a learned similarity function and a Parzen-window interpretation of Bayes classification. Its multiclass extension achieves 79.97% accuracy on PathMNIST, 80.46% on CIFAR-10, and 58.49% on SleepEDF; in sepsis prediction, the cosine model reaches 0.848 AUROC while the Mahalanobis variant provides the best calibration with an ECE of 0.229.

Key insights

• The CLaI learning objective is algebraically equivalent to minimizing a kernel-alignment distance, linking neural representation learning with supervised kernel selection.
• Collaborative Inference can be understood as nonparametric Bayesian classification: kernel sums estimate class-conditional densities and empirical class frequencies estimate priors.
• Discrimination and calibration respond differently to the similarity choice: cosine similarity performed better on sepsis ranking and F1, whereas the Mahalanobis model produced lower expected calibration error and converged faster in the reported training comparison.

Summary by alphaXiv.