The Power of Max Pooling Layer

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F67985807%3A_____%2F25%3A00636921" target="_blank" >RIV/67985807:_____/25:00636921 - isvavai.cz</a>

Result on the web

<a href="https://doi.org/10.1007/978-3-032-04558-4_18" target="_blank" >https://doi.org/10.1007/978-3-032-04558-4_18</a>

DOI - Digital Object Identifier

<a href="http://dx.doi.org/10.1007/978-3-032-04558-4_18" target="_blank" >10.1007/978-3-032-04558-4_18</a>

Result language

angličtina

Original language name

The Power of Max Pooling Layer

Original language description

Max pooling layers are the basic building blocks of convolutional neural networks. The theoretical characterization of their computational power is therefore a question of central interest. This paper deals with the representability of the max pooling layer by neural networks (NNs) employing the ReLU activation function. We provide two upper bounds on the size (number of ReLU neurons) and depth (number of layers) of the NNs that implement the maximum MAXn of n nonnegative numbers. We show that the MAXn function can be computed either by a NN of size n and logarithmic depth, or by a NN of quadratic size and constant depth for bounded input numbers of limited precision, where the constant depth depends on the magnitude of the weights. As a lower bound, we prove that no NN of depth 2 can compute the maximum of more than two nonnegative numbers. This confirms that the max pooling layer cannot be replaced by just two convolutional layer

Czech name

—

Czech description

—

Type

D - Article in proceedings

CEP classification

—

OECD FORD branch

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Project

<a href="/en/project/GA25-15490S" target="_blank" >GA25-15490S: LEDNeCo: Low Energy Deep Neurocomputing</a><br>

Continuities

I - Institucionalni podpora na dlouhodoby koncepcni rozvoj vyzkumne organizace

Publication year

2025

Confidentiality

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Article name in the collection

Artificial Neural Networks and Machine Learning – ICANN 2025. Proceedings, Part I

ISBN

978-3-032-04557-7

ISSN

0302-9743

e-ISSN

—

Number of pages

13

Pages from-to

221-233

Publisher name

Springer

Place of publication

Cham

Event location

Kaunas

Event date

Sep 9, 2025

Type of event by nationality

EUR - Evropská akce

UT code for WoS article

—