kth.sePublications
Change search
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf
Extrapolation of polynomial nets and their generalization guarantees
KTH, School of Electrical Engineering and Computer Science (EECS).
2022 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE creditsStudent thesis
Abstract [en]

Polynomial neural networks (NNs-Hp) have recently demonstrated high expressivity and efficiency across several tasks. However, a theoretical explanation toward such success is still unclear, especially when compared to the classical neural networks. Neural tangent kernel (NTK) is a powerful tool to analyze the training dynamics of neural networks and their generalization bounds. The study on NTK has been devoted to typical neural network architectures, but is incomplete for NNs-Hp. In this work, we derive the finite-width NTK formulation for NNs-Hp, and prove their equivalence to the kernel regression predictor with the associated NTK, which expands the application scope of NTK. Based on our results, we elucidate the difference of NNs-Hp over standard neural networks with respect to extrapolation and spectral bias. Our two key insights are that when compared to standard neural networks, a) NNs-Hp are able to fit more complicated functions in the extrapolation region; and b) NNs-Hp admit a slower eigenvalue decay of the respective NTK. Our empirical results provide a good justification for a deeper understanding of NNs-Hp

Abstract [sv]

Polynomiska neurala nätverk (NNs-Hp) har nyligen visat hög uttrycksförmåga och effektivitet över flera uppgifter. En teoretisk förklaring till sådan framgång är dock fortfarande oklar, särskilt jämfört med de klassiska neurala nätverken. Neurala tangentkärnor (NTK) är ett kraftfullt verktyg för att analysera träningsdynamiken i neurala nätverk och deras generaliseringsgränser. Studien om NTK har ägnats åt typiska neurala nätverksarkitekturer, men är ofullständig för NNs-Hp. I detta arbete härleder vi NTK-formuleringen med ändlig bredd för NNs-Hp och bevisar deras likvärdighet med kärnregressionsprediktorn med den associerade NTK, vilket utökar tillämpningsomfånget för NTK. Baserat på våra resultat belyser vi skillnaden mellan NNs-Hp jämfört med standardneurala nätverk med avseende på extrapolering och spektral bias. Våra två viktiga insikter är att jämfört med vanliga neurala nätverk, a) NNs-Hp kan passa mer komplicerade funktioner i extrapolationsregionen; och b) NNs-Hp medger en långsammare egenvärdesavklingning av respektive NTK. Våra empiriska resultat ger en bra motivering för en djupare förståelse av NNs-Hp.

Place, publisher, year, edition, pages
2022. , p. 35
Series
TRITA-EECS-EX ; 2022:229
Keywords [en]
Polynomial neural networks, Neural tangent kernel, Extrapolation
Keywords [sv]
Polynomial neural networks, Neural tangent kernel, Extrapolation
National Category
Computer and Information Sciences
Identifiers
URN: urn:nbn:se:kth:diva-318819OAI: oai:DiVA.org:kth-318819DiVA, id: diva2:1698082
External cooperation
Swiss Federal Institute of Technology Lausanne
Subject / course
Mechanical Engineering
Educational program
Master of Science - Machine Learning
Supervisors
Examiners
Available from: 2022-09-23 Created: 2022-09-22 Last updated: 2022-09-23Bibliographically approved

Open Access in DiVA

fulltext(2704 kB)247 downloads
File information
File name FULLTEXT01.pdfFile size 2704 kBChecksum SHA-512
e67462f00c08374ea6317fe5a80cbf8c67279feb965b3b1f5471cc9ec2cfe3328ea2fc0cab491ea63670edd2bcdc0230aa285728f8e812761b16ab90f79af0a4
Type fulltextMimetype application/pdf

By organisation
School of Electrical Engineering and Computer Science (EECS)
Computer and Information Sciences

Search outside of DiVA

GoogleGoogle Scholar
Total: 247 downloads
The number of downloads is the sum of all downloads of full texts. It may include eg previous versions that are now no longer available

urn-nbn

Altmetric score

urn-nbn
Total: 401 hits
CiteExportLink to record
Permanent link

Direct link
Cite
Citation style
  • apa
  • ieee
  • modern-language-association-8th-edition
  • vancouver
  • Other style
More styles
Language
  • de-DE
  • en-GB
  • en-US
  • fi-FI
  • nn-NO
  • nn-NB
  • sv-SE
  • Other locale
More languages
Output format
  • html
  • text
  • asciidoc
  • rtf