State-space representations of deep neural networks

Michael Hauser, Sean Gunn, Samer Saab, Asok Ray

Research output: Contribution to journalArticle

Abstract

This letter deals with neural networks as dynamical systems governed by finite difference equations. It shows that the introduction of k-many skip connections into network architectures, such as residual networks and additive dense networks, defines kth order dynamical equations on the layer-wise transformations. Closed-form solutions for the state-space representations of general kth order additive dense networks, where the concatenation operation is replaced by addition, as well as kth order smooth networks, are found. The developed provision endows deep neural networks with an algebraic structure. Furthermore, it is shown that imposing kth order smoothness on network architectures with d-many nodes per layer increases the state-space dimension by a multiple of k, and so the effective embedding dimension of the data manifold by the neural network is k·d-many dimensions. It follows that network architectures of these types reduce the number of parameters needed to maintain the same embedding dimension by a factor of k2 when compared to an equivalent first-order, residual network. Numerical simulations and experiments on CIFAR10, SVHN, and MNIST have been conducted to help understand the developed theory and efficacy of the proposed concepts.

Original languageEnglish (US)
Pages (from-to)538-554
Number of pages17
JournalNeural computation
Volume31
Issue number3
DOIs
StatePublished - Mar 1 2019

Fingerprint

Neural Networks
Equations
Layer
Simulation
Letters
Dynamical Systems
Efficacy
Algebra
Experiment

All Science Journal Classification (ASJC) codes

  • Arts and Humanities (miscellaneous)
  • Cognitive Neuroscience

Cite this

Hauser, Michael ; Gunn, Sean ; Saab, Samer ; Ray, Asok. / State-space representations of deep neural networks. In: Neural computation. 2019 ; Vol. 31, No. 3. pp. 538-554.
@article{d590e03fb3d94b0392de0ec8c6d3df81,
title = "State-space representations of deep neural networks",
abstract = "This letter deals with neural networks as dynamical systems governed by finite difference equations. It shows that the introduction of k-many skip connections into network architectures, such as residual networks and additive dense networks, defines kth order dynamical equations on the layer-wise transformations. Closed-form solutions for the state-space representations of general kth order additive dense networks, where the concatenation operation is replaced by addition, as well as kth order smooth networks, are found. The developed provision endows deep neural networks with an algebraic structure. Furthermore, it is shown that imposing kth order smoothness on network architectures with d-many nodes per layer increases the state-space dimension by a multiple of k, and so the effective embedding dimension of the data manifold by the neural network is k·d-many dimensions. It follows that network architectures of these types reduce the number of parameters needed to maintain the same embedding dimension by a factor of k2 when compared to an equivalent first-order, residual network. Numerical simulations and experiments on CIFAR10, SVHN, and MNIST have been conducted to help understand the developed theory and efficacy of the proposed concepts.",
author = "Michael Hauser and Sean Gunn and Samer Saab and Asok Ray",
year = "2019",
month = "3",
day = "1",
doi = "10.1162/neco_a_01165",
language = "English (US)",
volume = "31",
pages = "538--554",
journal = "Neural Computation",
issn = "0899-7667",
publisher = "MIT Press Journals",
number = "3",

}

State-space representations of deep neural networks. / Hauser, Michael; Gunn, Sean; Saab, Samer; Ray, Asok.

In: Neural computation, Vol. 31, No. 3, 01.03.2019, p. 538-554.

Research output: Contribution to journalArticle

TY - JOUR

T1 - State-space representations of deep neural networks

AU - Hauser, Michael

AU - Gunn, Sean

AU - Saab, Samer

AU - Ray, Asok

PY - 2019/3/1

Y1 - 2019/3/1

N2 - This letter deals with neural networks as dynamical systems governed by finite difference equations. It shows that the introduction of k-many skip connections into network architectures, such as residual networks and additive dense networks, defines kth order dynamical equations on the layer-wise transformations. Closed-form solutions for the state-space representations of general kth order additive dense networks, where the concatenation operation is replaced by addition, as well as kth order smooth networks, are found. The developed provision endows deep neural networks with an algebraic structure. Furthermore, it is shown that imposing kth order smoothness on network architectures with d-many nodes per layer increases the state-space dimension by a multiple of k, and so the effective embedding dimension of the data manifold by the neural network is k·d-many dimensions. It follows that network architectures of these types reduce the number of parameters needed to maintain the same embedding dimension by a factor of k2 when compared to an equivalent first-order, residual network. Numerical simulations and experiments on CIFAR10, SVHN, and MNIST have been conducted to help understand the developed theory and efficacy of the proposed concepts.

AB - This letter deals with neural networks as dynamical systems governed by finite difference equations. It shows that the introduction of k-many skip connections into network architectures, such as residual networks and additive dense networks, defines kth order dynamical equations on the layer-wise transformations. Closed-form solutions for the state-space representations of general kth order additive dense networks, where the concatenation operation is replaced by addition, as well as kth order smooth networks, are found. The developed provision endows deep neural networks with an algebraic structure. Furthermore, it is shown that imposing kth order smoothness on network architectures with d-many nodes per layer increases the state-space dimension by a multiple of k, and so the effective embedding dimension of the data manifold by the neural network is k·d-many dimensions. It follows that network architectures of these types reduce the number of parameters needed to maintain the same embedding dimension by a factor of k2 when compared to an equivalent first-order, residual network. Numerical simulations and experiments on CIFAR10, SVHN, and MNIST have been conducted to help understand the developed theory and efficacy of the proposed concepts.

UR - http://www.scopus.com/inward/record.url?scp=85061596492&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85061596492&partnerID=8YFLogxK

U2 - 10.1162/neco_a_01165

DO - 10.1162/neco_a_01165

M3 - Article

C2 - 30645180

AN - SCOPUS:85061596492

VL - 31

SP - 538

EP - 554

JO - Neural Computation

JF - Neural Computation

SN - 0899-7667

IS - 3

ER -