Laws of Learning Dynamics and the Core of Learners

Inkee Jung, Siu Cheong Lau · 2026

Read Laws of Learning Dynamics and the Core of Learners by Inkee Jung, Siu Cheong Lau (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-05026.txt.

Description

Inkee Jung and Siu Cheong Lau propose basic laws for how learning itself unfolds in this 2026 machine learning paper. Their framework states a conservation law and a law that total entropy decreases during learning. From those laws they build a method for ensembles that keep learning over a lifetime, rather than training once and freezing.

The test bed is defense against transfer-based adversarial attacks on the CIFAR-10 image dataset. Such attacks craft misleading inputs on one model and reuse them against another. The authors' entropy-based ensemble, nicknamed a logifold, immunizes a model by combining versions trained on clean and attacked samples. Compared with simply averaging those models, the logifold scored higher accuracy in most test cases, with especially large gains when the attacks were strong.

Catalog record

CREATORInkee Jung, Siu Cheong Lau
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2602-05026

Browse the full library · About the Tower of Babel