Scientists develop new mathematical mannequin for switch studying in neural networks

Could 7, 2025

The GIST Editors' notes

This text has been reviewed in accordance with Science X's editorial course of and insurance policies. Editors have highlighted the next attributes whereas guaranteeing the content material's credibility:

fact-checked

peer-reviewed publication

trusted supply

proofread

Scientists develop new mathematical mannequin for switch studying in neural networks

Alessandro Ingrosso, researcher on the Donders Institute for Neuroscience, has developed a brand new mathematical methodology in collaboration with colleagues in two Italian analysis establishments, which permits the prediction of the effectiveness of switch studying in neural networks.

The paper is printed within the journal Bodily Evaluation Letters.

The issue of restricted knowledge

AI methods are usually skilled with massive quantities of information. Contemplate a picture recognition system skilled to establish a canine in an image. By coaching this community with massive portions of canine images, the community learns to establish a canine with excessive accuracy in new pictures.

Nevertheless, for some purposes, ample coaching knowledge is just not accessible. "Within the medical sector, for instance, when diagnosing most cancers by way of ultrasound, there aren't sufficient examples accessible to correctly practice a neural community," explains Ingrosso.

"This results in 'overfitting,' the place the community fails at generalizing on new, unseen instances. The variety of false positives and false negatives can then turn into very massive."

To resolve this downside, the researchers centered on "switch studying," a method the place data from a community skilled on a big dataset (the "supply") is transferred to a brand new community (the "goal") that’s skilled on restricted knowledge.

"We studied the only mannequin for this switch studying method and developed a mathematical principle for networks with one hidden layer," says Ingrosso.

Modern mixture of analytical strategies

On this principle, the researchers mixed two totally different analytical strategies: the just lately developed "Kernel Renormalization" method and the classical "Franz-Parisi" formalism from the speculation of Spin Glasses. Merging such instruments makes it potential to work with particular, actual datasets as a substitute of statistical fashions or approximations.

Ingrosso provides, "Our new methodology can immediately and precisely predict how efficient the goal community will probably be in generalizing knowledge when it adopts data from the supply community."

This analysis gives necessary new insights for the efficient coaching of AI methods in domains the place knowledge is scarce, equivalent to medical diagnostics and different specialised purposes.

Extra info: Alessandro Ingrosso et al, Statistical Mechanics of Switch Studying in Absolutely Linked Networks within the Proportional Restrict, Bodily Evaluation Letters (2025). DOI: 10.1103/PhysRevLett.134.177301. On arXiv: DOI: 10.48550/arxiv.2407.07168

Journal info: Physical Review Letters , arXiv Offered by Radboud College Quotation: Scientists develop new mathematical mannequin for switch studying in neural networks (2025, Could 7) retrieved 7 Could 2025 from https://techxplore.com/information/2025-05-scientists-mathematical-neural-networks.html This doc is topic to copyright. Aside from any honest dealing for the aim of personal research or analysis, no half could also be reproduced with out the written permission. The content material is offered for info functions solely.

Discover additional

Radiologists might quickly be utilizing AI to detect mind tumors 34 shares