Fisher information and the geometry of memorization in neural networks
Abstract
A fruitful approach towards understanding generalization in machine learning involves characterizing how models separate generalizable structure from idiosyncratic or even corrupted features of training data. Fisher information has emerged as a diagnostic tool for this separation, leveraging loss geometry to distinguish generalization from memorization. Here we investigate this connection in vision models trained on CIFAR-10 with controlled noise that can be learned only through memorization. We show that projecting model weights onto the top Fisher eigenvectors decouples general task performance from noisy sample memorization, and we investigate how this decoupling presents differentially across layers. We show that the same phenomena are present in small transformers trained to solve a synthetic in-context learning task, and are reproduced in closed form in a simple, analytically solvable linear model. Our results reveal that Fisher information captures aspects of the geometric structure underlying how neural networks allocate capacity between generalization and memorization.
Related
- Fisher-Information Geometry Linking Schrodinger Dynamics, Bipartite Correlations, and Decoherence
- Informative Fisher Modes in Neural Networks as a Result of Explicit Symmetry Breaking
- Emergent Loop Geometry in Gauge-Equivariant Graph Neural Networks
- Inferring Nuclear Radial Parameters from Initial-State Information in Heavy-Ion Collisions Using Deep Neural Networks
- From Vacuum Geometry to Solution-Phase Electronic Properties and Thermodynamics via Equivariant Neural Networks
BibTeX
@misc{bernstein2026fisher,
title = {{Fisher information and the geometry of memorization in neural networks}},
author = {Daniel T. Bernstein and Chase Goddard and Luca Di Carlo and Vudtiwat Ngampruetikorn and David J. Schwab},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.16161},
eprint = {2610.16161},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.16161},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Fisher information and the geometry of memorization in neural networks
AU - Daniel T. Bernstein
AU - Chase Goddard
AU - Luca Di Carlo
AU - Vudtiwat Ngampruetikorn
AU - David J. Schwab
PY - 2026
DA - 2026/10/12
PB - arXiv
DO - 10.48550/arxiv.2610.16161
UR - https://arxiv.org/abs/2610.16161
AB - A fruitful approach towards understanding generalization in machine learning involves characterizing how models separate generalizable structure from idiosyncratic or even corrupted features of training data. Fisher information has emerged as a diagnostic tool for this separation, leveraging loss geometry to distinguish generalization from memorization. Here we investigate this connection in vision models trained on CIFAR-10 with controlled noise that can be learned only through memorization. We show that projecting model weights onto the top Fisher eigenvectors decouples general task performance from noisy sample memorization, and we investigate how this decoupling presents differentially across layers. We show that the same phenomena are present in small transformers trained to solve a synthetic in-context learning task, and are reproduced in closed form in a simple, analytically solvable linear model. Our results reveal that Fisher information captures aspects of the geometric structure underlying how neural networks allocate capacity between generalization and memorization.
ER -