On the Depth of Monotone ReLU Neural Networks and ICNNs

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Bakaev, Egor, Brunck, Florestan, Hertrich, Christoph, Reichman, Daniel, Yehudayoff, Amir
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908356519657472
author Bakaev, Egor
Brunck, Florestan
Hertrich, Christoph
Reichman, Daniel
Yehudayoff, Amir
author_facet Bakaev, Egor
Brunck, Florestan
Hertrich, Christoph
Reichman, Daniel
Yehudayoff, Amir
contents We study two models of ReLU neural networks: monotone networks (ReLU$^+$) and input convex neural networks (ICNN). Our focus is on expressivity, mostly in terms of depth, and we prove the following lower bounds. For the maximum function MAX$_n$ computing the maximum of $n$ real numbers, we show that ReLU$^+$ networks cannot compute MAX$_n$, or even approximate it. We prove a sharp $n$ lower bound on the ICNN depth complexity of MAX$_n$. We also prove depth separations between ReLU networks and ICNNs; for every $k$, there is a depth-2 ReLU network of size $O(k^2)$ that cannot be simulated by a depth-$k$ ICNN. The proofs are based on deep connections between neural networks and polyhedral geometry, and also use isoperimetric properties of triangulations.
format Preprint
id arxiv_https___arxiv_org_abs_2505_06169
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle On the Depth of Monotone ReLU Neural Networks and ICNNs
Bakaev, Egor
Brunck, Florestan
Hertrich, Christoph
Reichman, Daniel
Yehudayoff, Amir
Machine Learning
Discrete Mathematics
Neural and Evolutionary Computing
Combinatorics
We study two models of ReLU neural networks: monotone networks (ReLU$^+$) and input convex neural networks (ICNN). Our focus is on expressivity, mostly in terms of depth, and we prove the following lower bounds. For the maximum function MAX$_n$ computing the maximum of $n$ real numbers, we show that ReLU$^+$ networks cannot compute MAX$_n$, or even approximate it. We prove a sharp $n$ lower bound on the ICNN depth complexity of MAX$_n$. We also prove depth separations between ReLU networks and ICNNs; for every $k$, there is a depth-2 ReLU network of size $O(k^2)$ that cannot be simulated by a depth-$k$ ICNN. The proofs are based on deep connections between neural networks and polyhedral geometry, and also use isoperimetric properties of triangulations.
title On the Depth of Monotone ReLU Neural Networks and ICNNs
topic Machine Learning
Discrete Mathematics
Neural and Evolutionary Computing
Combinatorics
url https://arxiv.org/abs/2505.06169