Optimal errors and phase transitions in high-dimensional generalized linear models

Generalized linear models (GLMs) are used in high-dimensional machine learning, statistics, communications, and signal processing. In this paper we analyze GLMs when the data matrix is random, as relevant in problems such as compressed sensing, error-correcting codes, or benchmark models in neural n...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Proceedings of the National Academy of Sciences - PNAS Ročník 116; číslo 12; s. 5451
Hlavní autoři: Barbier, Jean, Krzakala, Florent, Macris, Nicolas, Miolane, Léo, Zdeborová, Lenka
Médium: Journal Article
Jazyk:angličtina
Vydáno: United States 19.03.2019
Témata:
ISSN:1091-6490, 1091-6490
On-line přístup:Zjistit podrobnosti o přístupu
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Abstract Generalized linear models (GLMs) are used in high-dimensional machine learning, statistics, communications, and signal processing. In this paper we analyze GLMs when the data matrix is random, as relevant in problems such as compressed sensing, error-correcting codes, or benchmark models in neural networks. We evaluate the mutual information (or "free entropy") from which we deduce the Bayes-optimal estimation and generalization errors. Our analysis applies to the high-dimensional limit where both the number of samples and the dimension are large and their ratio is fixed. Nonrigorous predictions for the optimal errors existed for special cases of GLMs, e.g., for the perceptron, in the field of statistical physics based on the so-called replica method. Our present paper rigorously establishes those decades-old conjectures and brings forward their algorithmic interpretation in terms of performance of the generalized approximate message-passing algorithm. Furthermore, we tightly characterize, for many learning problems, regions of parameters for which this algorithm achieves the optimal performance and locate the associated sharp phase transitions separating learnable and nonlearnable regions. We believe that this random version of GLMs can serve as a challenging benchmark for multipurpose algorithms.
AbstractList Generalized linear models (GLMs) are used in high-dimensional machine learning, statistics, communications, and signal processing. In this paper we analyze GLMs when the data matrix is random, as relevant in problems such as compressed sensing, error-correcting codes, or benchmark models in neural networks. We evaluate the mutual information (or "free entropy") from which we deduce the Bayes-optimal estimation and generalization errors. Our analysis applies to the high-dimensional limit where both the number of samples and the dimension are large and their ratio is fixed. Nonrigorous predictions for the optimal errors existed for special cases of GLMs, e.g., for the perceptron, in the field of statistical physics based on the so-called replica method. Our present paper rigorously establishes those decades-old conjectures and brings forward their algorithmic interpretation in terms of performance of the generalized approximate message-passing algorithm. Furthermore, we tightly characterize, for many learning problems, regions of parameters for which this algorithm achieves the optimal performance and locate the associated sharp phase transitions separating learnable and nonlearnable regions. We believe that this random version of GLMs can serve as a challenging benchmark for multipurpose algorithms.Generalized linear models (GLMs) are used in high-dimensional machine learning, statistics, communications, and signal processing. In this paper we analyze GLMs when the data matrix is random, as relevant in problems such as compressed sensing, error-correcting codes, or benchmark models in neural networks. We evaluate the mutual information (or "free entropy") from which we deduce the Bayes-optimal estimation and generalization errors. Our analysis applies to the high-dimensional limit where both the number of samples and the dimension are large and their ratio is fixed. Nonrigorous predictions for the optimal errors existed for special cases of GLMs, e.g., for the perceptron, in the field of statistical physics based on the so-called replica method. Our present paper rigorously establishes those decades-old conjectures and brings forward their algorithmic interpretation in terms of performance of the generalized approximate message-passing algorithm. Furthermore, we tightly characterize, for many learning problems, regions of parameters for which this algorithm achieves the optimal performance and locate the associated sharp phase transitions separating learnable and nonlearnable regions. We believe that this random version of GLMs can serve as a challenging benchmark for multipurpose algorithms.
Generalized linear models (GLMs) are used in high-dimensional machine learning, statistics, communications, and signal processing. In this paper we analyze GLMs when the data matrix is random, as relevant in problems such as compressed sensing, error-correcting codes, or benchmark models in neural networks. We evaluate the mutual information (or "free entropy") from which we deduce the Bayes-optimal estimation and generalization errors. Our analysis applies to the high-dimensional limit where both the number of samples and the dimension are large and their ratio is fixed. Nonrigorous predictions for the optimal errors existed for special cases of GLMs, e.g., for the perceptron, in the field of statistical physics based on the so-called replica method. Our present paper rigorously establishes those decades-old conjectures and brings forward their algorithmic interpretation in terms of performance of the generalized approximate message-passing algorithm. Furthermore, we tightly characterize, for many learning problems, regions of parameters for which this algorithm achieves the optimal performance and locate the associated sharp phase transitions separating learnable and nonlearnable regions. We believe that this random version of GLMs can serve as a challenging benchmark for multipurpose algorithms.
Author Miolane, Léo
Krzakala, Florent
Macris, Nicolas
Zdeborová, Lenka
Barbier, Jean
Author_xml – sequence: 1
  givenname: Jean
  surname: Barbier
  fullname: Barbier, Jean
  email: jbarbier@ictp.it, leo.miolane@gmail.com
  organization: Laboratoire de Physique de l'Ecole Normale Supérieure, Université Paris-Sciences-et-Lettres, Centre National de la Recherche Scientifique, Sorbonne Université, Université Paris-Diderot, Sorbonne Paris Cité, 75005 Paris, France
– sequence: 2
  givenname: Florent
  surname: Krzakala
  fullname: Krzakala, Florent
  organization: Laboratoire de Physique de l'Ecole Normale Supérieure, Université Paris-Sciences-et-Lettres, Centre National de la Recherche Scientifique, Sorbonne Université, Université Paris-Diderot, Sorbonne Paris Cité, 75005 Paris, France
– sequence: 3
  givenname: Nicolas
  orcidid: 0000-0003-2189-7411
  surname: Macris
  fullname: Macris, Nicolas
  organization: Communication Theory Laboratory, School of Computer and Communication Sciences, Ecole Polytechnique Fédérale de Lausanne, CH-1015 Lausanne, Switzerland
– sequence: 4
  givenname: Léo
  surname: Miolane
  fullname: Miolane, Léo
  email: jbarbier@ictp.it, leo.miolane@gmail.com
  organization: Département d'Informatique de l'Ecole Normale Supérieure, Université Paris-Sciences-et-Lettres, Centre National de la Recherche Scientifique, Inria, 75005 Paris, France; jbarbier@ictp.it leo.miolane@gmail.com
– sequence: 5
  givenname: Lenka
  surname: Zdeborová
  fullname: Zdeborová, Lenka
  organization: Institut de Physique Théorique, Centre National de la Recherche Scientifique et Commissariat à l'Energie Atomique, Université Paris-Saclay, 91191 Gif-sur-Yvette, France
BackLink https://www.ncbi.nlm.nih.gov/pubmed/30824595$$D View this record in MEDLINE/PubMed
BookMark eNpNUEtLxDAYDLLiPvTsTXL00vVL0rTJURZfsLAgei7Z5utuJE1r0h7011twBU8zDDMDM0syC11AQq4ZrBmU4q4PJq2ZAl6CZKw4IwsGmmVFrmH2j8_JMqUPANBSwQWZC1A8l1ouyOuuH1xrPMUYu5ioCZb2R5OQDtGE5AbXhURdoEd3OGbWtTiJXZgCBwwYjXffaKl3AU2kbWfRp0ty3hif8OqEK_L--PC2ec62u6eXzf02q6VkQ4ZoG1NjaaG2VhrYcy2KOm8EFEogy4UtZdE0RpSWa62U1ZCLGnXJjFbF3vIVuf3t7WP3OWIaqtalGr03AbsxVZypUgouQUzWm5N13Ldoqz5Om-NX9fcD_wG8cGL3
CitedBy_id crossref_primary_10_1007_s10955_020_02495_2
crossref_primary_10_1088_1742_5468_ab43d2
crossref_primary_10_1007_s00440_021_01092_y
crossref_primary_10_1214_20_AOS2038
crossref_primary_10_1088_1751_8121_acb530
crossref_primary_10_1007_s10208_021_09531_x
crossref_primary_10_1109_TIT_2024_3481956
crossref_primary_10_1109_TCOMM_2021_3094509
crossref_primary_10_1088_1742_5468_ac9828
crossref_primary_10_1088_2632_2153_ace60f
crossref_primary_10_1088_1751_8121_ab8416
crossref_primary_10_1109_TCOMM_2024_3395337
crossref_primary_10_1109_TIT_2024_3415654
crossref_primary_10_1109_TIT_2022_3225802
crossref_primary_10_1109_TIT_2020_2990880
crossref_primary_10_1073_pnas_2311810121
crossref_primary_10_1103_PhysRevResearch_7_013081
crossref_primary_10_1214_23_AOS2309
crossref_primary_10_1109_TIT_2023_3307553
crossref_primary_10_1214_24_AOS2449
crossref_primary_10_1214_23_AOS2301
crossref_primary_10_1214_21_AOS2100
crossref_primary_10_1007_s00220_022_04387_w
crossref_primary_10_1088_1742_5468_ac59ac
crossref_primary_10_1109_TIT_2025_3560070
crossref_primary_10_1103_PhysRevX_15_021085
crossref_primary_10_1214_25_AOS2489
crossref_primary_10_1038_s41467_023_36657_z
crossref_primary_10_1109_TIT_2023_3321575
crossref_primary_10_1088_1751_8121_ab2735
crossref_primary_10_1109_TIT_2022_3182018
crossref_primary_10_1109_TIT_2024_3396472
crossref_primary_10_1103_PhysRevResearch_4_013201
crossref_primary_10_1109_TIT_2020_3015173
crossref_primary_10_1016_j_physa_2022_128154
crossref_primary_10_1016_j_physleta_2021_127756
crossref_primary_10_1088_1742_5468_ac3ae6
crossref_primary_10_1088_1742_5468_adfa75
crossref_primary_10_1088_1742_5468_ada696
crossref_primary_10_1088_1742_5468_adde3e
crossref_primary_10_1109_TNNLS_2021_3050422
crossref_primary_10_1111_stan_12274
crossref_primary_10_1088_1751_8121_ab59ef
crossref_primary_10_1088_1751_8121_ab8ff4
crossref_primary_10_1109_TIT_2022_3163342
crossref_primary_10_3150_21_BEJ1401
crossref_primary_10_1088_1742_5468_ababff
crossref_primary_10_1088_2632_2153_acb428
crossref_primary_10_1093_imaiai_iaaf019
crossref_primary_10_1016_j_acha_2024_101736
crossref_primary_10_1088_1742_5468_ab3430
crossref_primary_10_1088_1742_5468_ade137
crossref_primary_10_1073_pnas_2302028120
crossref_primary_10_1103_PhysRevX_10_011057
crossref_primary_10_1103_PhysRevE_106_025304
crossref_primary_10_1109_TIT_2023_3299490
crossref_primary_10_1088_1742_5468_ab4bbb
crossref_primary_10_1088_1742_5468_ac7e4c
crossref_primary_10_1073_pnas_2403682121
crossref_primary_10_1088_1742_5468_ac2edd
crossref_primary_10_1109_TWC_2021_3139034
crossref_primary_10_1093_imaiai_iaaf024
crossref_primary_10_1088_2632_2153_ac0615
crossref_primary_10_1109_TIT_2024_3455228
crossref_primary_10_1103_PhysRevX_10_041044
crossref_primary_10_1088_1742_5468_adb1d6
crossref_primary_10_1088_1742_5468_abc61d
crossref_primary_10_1007_s10955_019_02470_6
crossref_primary_10_1093_jrsssb_qkae039
crossref_primary_10_1088_1751_8121_ad2c26
crossref_primary_10_1088_1742_5468_ad65e6
crossref_primary_10_1088_1751_8121_aba028
crossref_primary_10_1109_TIT_2023_3311408
crossref_primary_10_1088_2632_2153_acd749
crossref_primary_10_1093_imaiai_iaae024
crossref_primary_10_1088_1742_5468_ad65e1
crossref_primary_10_1109_TSP_2020_2974711
crossref_primary_10_1088_1742_5468_ac3a75
crossref_primary_10_1093_imaiai_iaab027
crossref_primary_10_1109_TIT_2024_3449321
crossref_primary_10_1103_l4p2_vrxt
crossref_primary_10_1088_2632_2153_abfbbb
crossref_primary_10_1007_s00220_022_04630_4
crossref_primary_10_1214_25_AAP2186
crossref_primary_10_1088_2632_2153_ac7d3b
crossref_primary_10_1109_TIT_2024_3486685
crossref_primary_10_1088_1742_5468_ac9cc8
crossref_primary_10_1214_25_STS993
crossref_primary_10_1109_TIT_2023_3325187
crossref_primary_10_1088_1742_5468_add1ce
crossref_primary_10_1088_1742_5468_ab7123
crossref_primary_10_1109_TIT_2024_3458953
crossref_primary_10_1109_TIT_2025_3535923
crossref_primary_10_1080_03610926_2024_2322616
crossref_primary_10_1109_TIT_2020_3033985
crossref_primary_10_1038_s42005_024_01870_9
crossref_primary_10_1214_21_AIHP1183
crossref_primary_10_1109_TCOMM_2023_3294960
crossref_primary_10_1214_24_AAP2120
crossref_primary_10_1214_22_AAP1874
ContentType Journal Article
Copyright Copyright © 2019 the Author(s). Published by PNAS.
Copyright_xml – notice: Copyright © 2019 the Author(s). Published by PNAS.
DBID NPM
7X8
DOI 10.1073/pnas.1802705116
DatabaseName PubMed
MEDLINE - Academic
DatabaseTitle PubMed
MEDLINE - Academic
DatabaseTitleList MEDLINE - Academic
PubMed
Database_xml – sequence: 1
  dbid: NPM
  name: PubMed
  url: http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?db=PubMed
  sourceTypes: Index Database
– sequence: 2
  dbid: 7X8
  name: MEDLINE - Academic
  url: https://search.proquest.com/medline
  sourceTypes: Aggregation Database
DeliveryMethod no_fulltext_linktorsrc
Discipline Sciences (General)
EISSN 1091-6490
ExternalDocumentID 30824595
Genre Research Support, Non-U.S. Gov't
Journal Article
GroupedDBID ---
-DZ
-~X
.55
0R~
123
29P
2AX
2FS
2WC
4.4
53G
5RE
5VS
85S
AACGO
AAFWJ
AANCE
ABBHK
ABOCM
ABPLY
ABPPZ
ABTLG
ABXSQ
ABZEH
ACGOD
ACHIC
ACIWK
ACNCT
ACPRK
ADQXQ
ADULT
AENEX
AEUPB
AEXZC
AFFNX
AFOSN
AFRAH
ALMA_UNASSIGNED_HOLDINGS
AQVQM
BKOMP
CS3
D0L
DCCCD
DIK
DU5
E3Z
EBS
EJD
F5P
FRP
GX1
H13
HH5
HYE
IPSME
JAAYA
JBMMH
JENOY
JHFFW
JKQEH
JLS
JLXEF
JPM
JSG
JST
KQ8
L7B
LU7
N9A
NPM
N~3
O9-
OK1
PNE
PQQKQ
R.V
RHI
RNA
RNS
RPM
RXW
SA0
SJN
TAE
TN5
UKR
W8F
WH7
WOQ
WOW
X7M
XSW
Y6R
YBH
YKV
YSK
ZCA
~02
~KM
7X8
ID FETCH-LOGICAL-c551t-eedface7d0cdd5a0b2936c4f30683e143d756ffa37d29988d9043ce971a986bd2
IEDL.DBID 7X8
ISICitedReferencesCount 177
ISICitedReferencesURI http://www.webofscience.com/api/gateway?GWVersion=2&SrcApp=Summon&SrcAuth=ProQuest&DestLinkType=CitingArticles&DestApp=WOS_CPL&KeyUT=000461679000043&url=https%3A%2F%2Fcvtisr.summon.serialssolutions.com%2F%23%21%2Fsearch%3Fho%3Df%26include.ft.matches%3Dt%26l%3Dnull%26q%3D
ISSN 1091-6490
IngestDate Sun Nov 09 10:32:19 EST 2025
Thu Apr 03 07:00:09 EDT 2025
IsDoiOpenAccess false
IsOpenAccess true
IsPeerReviewed true
IsScholarly true
Issue 12
Keywords generalized linear model
high-dimensional inference
approximate message-passing algorithm
perceptron
Bayesian inference
Language English
License Copyright © 2019 the Author(s). Published by PNAS.
LinkModel DirectLink
MergedId FETCHMERGED-LOGICAL-c551t-eedface7d0cdd5a0b2936c4f30683e143d756ffa37d29988d9043ce971a986bd2
Notes ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
ORCID 0000-0003-2189-7411
OpenAccessLink https://cea.hal.science/cea-01614258
PMID 30824595
PQID 2187532503
PQPubID 23479
ParticipantIDs proquest_miscellaneous_2187532503
pubmed_primary_30824595
PublicationCentury 2000
PublicationDate 2019-03-19
PublicationDateYYYYMMDD 2019-03-19
PublicationDate_xml – month: 03
  year: 2019
  text: 2019-03-19
  day: 19
PublicationDecade 2010
PublicationPlace United States
PublicationPlace_xml – name: United States
PublicationTitle Proceedings of the National Academy of Sciences - PNAS
PublicationTitleAlternate Proc Natl Acad Sci U S A
PublicationYear 2019
SSID ssj0009580
Score 2.6915462
Snippet Generalized linear models (GLMs) are used in high-dimensional machine learning, statistics, communications, and signal processing. In this paper we analyze...
SourceID proquest
pubmed
SourceType Aggregation Database
Index Database
StartPage 5451
Title Optimal errors and phase transitions in high-dimensional generalized linear models
URI https://www.ncbi.nlm.nih.gov/pubmed/30824595
https://www.proquest.com/docview/2187532503
Volume 116
WOSCitedRecordID wos000461679000043&url=https%3A%2F%2Fcvtisr.summon.serialssolutions.com%2F%23%21%2Fsearch%3Fho%3Df%26include.ft.matches%3Dt%26l%3Dnull%26q%3D
hasFullText
inHoldings 1
isFullTextHit
isPrint
link http://cvtisr.summon.serialssolutions.com/2.0.0/link/0/eLvHCXMwpV09T8MwELWAMrAA5bN8yUgMMBgSkvhjQghRMUCpEEjdKse-QAeSkBQGfj3nOBUsSEgsmZwoOj_fe7ZP7wg5AmfIIlPOMuQeFqPoZRq4YVkaSgSUTTM_07diMJCjkRq2B251W1Y5y4lNoraFcWfkZ0hFqKyRsKOL8o25rlHudrVtoTFPOhFKGYdqMZI_THeldyNQIeOxCmbWPiI6K3Ndnzr3M4GoDPnv-rLhmf7Kf_9wlSy3CpNeekh0yRzka6TbruGaHrdG0yfr5OEe88UrjoWqKqqa6tzS8gVpjU4dg_liLjrJqfM0Ztb1AfAeHvTZf2PyCZY6naor2rTUqTfIU__68eqGtT0WmEGtNGVIkZk2IGxgrE10kCL9cxNnuJOQEaCYsiLhWaYjYZG4pLQqiCMDSoRaSZ7a802ykBc5bBMKJgVQsVAgUeYEICOtQVqBO8iIy1j3yOEsbmPEsLuY0DkU7_X4O3I9suWDPy692cbY2enEiUp2_vD2LllCPaNciVio9kgnwxUM-2TRfEwndXXQgAOfg-HdF7K7xS0
linkProvider ProQuest
openUrl ctx_ver=Z39.88-2004&ctx_enc=info%3Aofi%2Fenc%3AUTF-8&rfr_id=info%3Asid%2Fsummon.serialssolutions.com&rft_val_fmt=info%3Aofi%2Ffmt%3Akev%3Amtx%3Ajournal&rft.genre=article&rft.atitle=Optimal+errors+and+phase+transitions+in+high-dimensional+generalized+linear+models&rft.jtitle=Proceedings+of+the+National+Academy+of+Sciences+-+PNAS&rft.au=Barbier%2C+Jean&rft.au=Krzakala%2C+Florent&rft.au=Macris%2C+Nicolas&rft.au=Miolane%2C+L%C3%A9o&rft.date=2019-03-19&rft.issn=1091-6490&rft.eissn=1091-6490&rft.volume=116&rft.issue=12&rft.spage=5451&rft_id=info:doi/10.1073%2Fpnas.1802705116&rft.externalDBID=NO_FULL_TEXT
thumbnail_l http://covers-cdn.summon.serialssolutions.com/index.aspx?isbn=/lc.gif&issn=1091-6490&client=summon
thumbnail_m http://covers-cdn.summon.serialssolutions.com/index.aspx?isbn=/mc.gif&issn=1091-6490&client=summon
thumbnail_s http://covers-cdn.summon.serialssolutions.com/index.aspx?isbn=/sc.gif&issn=1091-6490&client=summon