You are here

Proc Natl Acad Sci U S A DOI:10.1073/pnas.2115064119

Simple, fast, and flexible framework for matrix completion with infinite width neural networks.

Publication TypeJournal Article
Year of Publication2022
AuthorsRadhakrishnan, A, Stefanakis, G, Belkin, M, Uhler, C
JournalProc Natl Acad Sci U S A
Volume119
Issue16
Pagese2115064119
Date Published2022 Apr 19
ISSN1091-6490
KeywordsComputers, Image Processing, Computer-Assisted, Machine Learning, Neural Networks, Computer, Supervised Machine Learning
Abstract

SignificanceMatrix completion is a fundamental problem in machine learning that arises in various applications. We envision that our infinite width neural network framework for matrix completion will be easily deployable and produce strong baselines for a wide range of applications at limited computational costs. We demonstrate the flexibility of our framework through competitive results on virtual drug screening and image inpainting/reconstruction. Simplicity and speed are showcased by the fact that most results in this work require only a central processing unit and commodity hardware. Through its connection to semisupervised learning, our framework provides a principled approach for matrix completion that can be easily applied to problems well beyond those of image completion and virtual drug screening considered in this paper.

DOI10.1073/pnas.2115064119
Pubmed

https://www.ncbi.nlm.nih.gov/pubmed/35412891?dopt=Abstract

Alternate JournalProc Natl Acad Sci U S A
PubMed ID35412891
Grant ListDMS-1651995 / / National Science Foundation (NSF) /
N00014-17-1-2147 / / DOD | United States Navy | Office of Naval Research (ONR) /
N00014-18-1-2765 / / DOD | United States Navy | Office of Naval Research (ONR) /
MIT-IBM Watson AI Lab / / MIT-IBM Watson AI Lab /
Simons Investigator Award / / Simons Foundation /
IIS-1815697 / / National Science Foundation (NSF) /
DMS-2031883 / / National Science Foundation (NSF) /
814639 / / Simons Foundation /