Every Feedforward Neural Network Definable in an o-Minimal Structure Has Finite Sample Complexity

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Kratsios, Anastasis, Cousins, Gregory, Borde, Haitz Sáez de Ocáriz, Kim, Bum Jun, Brugiapaglia, Simone
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910200307384320
author Kratsios, Anastasis
Cousins, Gregory
Borde, Haitz Sáez de Ocáriz
Kim, Bum Jun
Brugiapaglia, Simone
author_facet Kratsios, Anastasis
Cousins, Gregory
Borde, Haitz Sáez de Ocáriz
Kim, Bum Jun
Brugiapaglia, Simone
contents We show that, in a precise sense, a broad class of feedforward neural networks learn (have finite sample complexity) in the PAC model: every fixed finite feedforward architecture whose layers are definable in an o-minimal structure has finite sample complexity in the agnostic PAC setting, even with unbounded parameters. This covers standard fixed-size MLPs, CNNs, GNNs, and transformers with fixed sequence length, together with the operations and layers typically used in such architectures, including linear projections, residual connections, attention mechanisms, pooling layers, normalization layers, and admissible positional encodings. Hence, distribution-free learnability for modern non-recurrent architectures is not an exceptional property of particular activations or architecture-specific VC arguments, but a consequence of tame feedforward computation. Our results reposition finite-sample PAC learnability as a baseline rather than a differentiator: they shift the focus of architectural comparison toward inductive biases, symmetries and geometric priors, scalability, and optimization behaviour.
format Preprint
id arxiv_https___arxiv_org_abs_2605_07097
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Every Feedforward Neural Network Definable in an o-Minimal Structure Has Finite Sample Complexity
Kratsios, Anastasis
Cousins, Gregory
Borde, Haitz Sáez de Ocáriz
Kim, Bum Jun
Brugiapaglia, Simone
Machine Learning
Neural and Evolutionary Computing
Logic
Statistics Theory
68Q32, 03C64, 03C98, 68T07, 62G05
We show that, in a precise sense, a broad class of feedforward neural networks learn (have finite sample complexity) in the PAC model: every fixed finite feedforward architecture whose layers are definable in an o-minimal structure has finite sample complexity in the agnostic PAC setting, even with unbounded parameters. This covers standard fixed-size MLPs, CNNs, GNNs, and transformers with fixed sequence length, together with the operations and layers typically used in such architectures, including linear projections, residual connections, attention mechanisms, pooling layers, normalization layers, and admissible positional encodings. Hence, distribution-free learnability for modern non-recurrent architectures is not an exceptional property of particular activations or architecture-specific VC arguments, but a consequence of tame feedforward computation. Our results reposition finite-sample PAC learnability as a baseline rather than a differentiator: they shift the focus of architectural comparison toward inductive biases, symmetries and geometric priors, scalability, and optimization behaviour.
title Every Feedforward Neural Network Definable in an o-Minimal Structure Has Finite Sample Complexity
topic Machine Learning
Neural and Evolutionary Computing
Logic
Statistics Theory
68Q32, 03C64, 03C98, 68T07, 62G05
url https://arxiv.org/abs/2605.07097