Algebraic Machine Learning for Small-to-Medium Datasets Is Competitive against Strong Standard Baselines
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.
Cite
@article{arxiv.2605.22155,
title = {Algebraic Machine Learning for Small-to-Medium Datasets Is Competitive against Strong Standard Baselines},
author = {David Mendez and Fernando Martin-Maroto and Gonzalo G. de Polavieja},
journal= {arXiv preprint arXiv:2605.22155},
year = {2026}
}
Comments
9 pages, 4 figures