Analytic function approximation by path norm regularized deep networks

Aleksandr Beknazaryan

doi:10.48550/arXiv.2104.02095

Analytic function approximation by path norm regularized deep networks

Aleksandr Beknazaryan

Mathematics of Operations Research

Research output: Working paper › Preprint › Academic

8 Downloads (Pure)

Abstract

We show that neural networks with absolute value activation function and with the path norm, the depth, the width and the network weights having logarithmic dependence on $1/\varepsilon$ can $\varepsilon$-approximate functions that are analytic on certain regions of $\mathbb{C}^d$.

Original language	English
Publisher	ArXiv.org
DOIs	https://doi.org/10.48550/arXiv.2104.02095
Publication status	Published - 5 Apr 2021

Keywords

stat.ML
cs.LG

Access to Document

10.48550/arXiv.2104.02095Licence: CC BY

2104.02095v4Final published version, 166 KBLicence: CC BY

Cite this

@techreport{a2f3748de94140d3a84246f06b955102,

title = "Analytic function approximation by path norm regularized deep networks",

abstract = " We show that neural networks with absolute value activation function and with the path norm, the depth, the width and the network weights having logarithmic dependence on $1/\varepsilon$ can $\varepsilon$-approximate functions that are analytic on certain regions of $\mathbb{C}^d$. ",

keywords = "stat.ML, cs.LG",

author = "Aleksandr Beknazaryan",

year = "2021",

month = apr,

day = "5",

doi = "10.48550/arXiv.2104.02095",

language = "English",

publisher = "ArXiv.org",

type = "WorkingPaper",

institution = "ArXiv.org",

}

TY - UNPB

T1 - Analytic function approximation by path norm regularized deep networks

AU - Beknazaryan, Aleksandr

PY - 2021/4/5

Y1 - 2021/4/5

N2 - We show that neural networks with absolute value activation function and with the path norm, the depth, the width and the network weights having logarithmic dependence on $1/\varepsilon$ can $\varepsilon$-approximate functions that are analytic on certain regions of $\mathbb{C}^d$.

AB - We show that neural networks with absolute value activation function and with the path norm, the depth, the width and the network weights having logarithmic dependence on $1/\varepsilon$ can $\varepsilon$-approximate functions that are analytic on certain regions of $\mathbb{C}^d$.

KW - stat.ML

KW - cs.LG

U2 - 10.48550/arXiv.2104.02095

DO - 10.48550/arXiv.2104.02095

M3 - Preprint

BT - Analytic function approximation by path norm regularized deep networks

PB - ArXiv.org

ER -

Analytic function approximation by path norm regularized deep networks

Abstract

Keywords

Access to Document

Fingerprint

Cite this