MMODELYST
Papers/Algebraic Machine Learning for Small-to-Medium Datasets Is Competitive against Strong Standard Baselines
PAP

Algebraic Machine Learning for Small-to-Medium Datasets Is Competitive against Strong Standard Baselines

May 21, 2026

arXiv
Abstract

Symbolic methods are generally not considered competitive with strong modern learners on realistic supervised tasks. We evaluate Algebraic Machine Learning (AML), a framework that learns through subdirect decomposition of algebraic structure rather than numerical optimization, against standard baselines on image and tabular classification across varying training-set sizes. We find that AML trained only on training data without using validation or cross-validation outperforms a family of cross-validated baseline methods including CNNs on small to medium image datasets (50--2000 training examples). On tabular datasets in the same size range, XGBoost is overall the best performing method, but AML is nonetheless comparable to methods incorporating task-specific biases such as LightGBM and random forests. AML achieves this competitive performance across two very different types of datasets using a generic algebraic inductive bias, rather than the modality-specific biases built into standard baselines like CNNs for images or XGBoost for tabular data, and requires no cross validation because it has no task-dependent hyperparameters to tune.

Select text to highlight · click a highlight to remove · saved in this browser only
Authors
David Mendez, Fernando Martin-Maroto, Gonzalo G. de Polavieja
Your notes (browser-local)
saved
arXiv:2605.22155