From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks

Biological and artificial neural networks develop internal representations that enable them to perform complex tasks. In artificial networks, the effectiveness of these models relies on their ability to build task specific representation, a process influenced by interactions among datasets, architec...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Dominé, Clémentine C. J, Anguita, Nicolas, Proca, Alexandra M, Braun, Lukas, Kunin, Daniel, Mediano, Pedro A. M, Saxe, Andrew M
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title
container_volume
creator Dominé, Clémentine C. J
Anguita, Nicolas
Proca, Alexandra M
Braun, Lukas
Kunin, Daniel
Mediano, Pedro A. M
Saxe, Andrew M
description Biological and artificial neural networks develop internal representations that enable them to perform complex tasks. In artificial networks, the effectiveness of these models relies on their ability to build task specific representation, a process influenced by interactions among datasets, architectures, initialization strategies, and optimization algorithms. Prior studies highlight that different initializations can place networks in either a lazy regime, where representations remain static, or a rich/feature learning regime, where representations evolve dynamically. Here, we examine how initialization influences learning dynamics in deep linear neural networks, deriving exact solutions for lambda-balanced initializations-defined by the relative scale of weights across layers. These solutions capture the evolution of representations and the Neural Tangent Kernel across the spectrum from the rich to the lazy regimes. Our findings deepen the theoretical understanding of the impact of weight initialization on learning regimes, with implications for continual learning, reversal learning, and transfer learning, relevant to both neuroscience and practical applications.
doi_str_mv 10.48550/arxiv.2409.14623
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2409_14623</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2409_14623</sourcerecordid><originalsourceid>FETCH-arxiv_primary_2409_146233</originalsourceid><addsrcrecordid>eNpjYJA0NNAzsTA1NdBPLKrILNMzMjGw1DM0MTMy5mRwcivKz1XwSayqVCjJVwjKTM6wUnCtSEwuUfBJTSzKy8xLV3CpzEvMzUwuVsjMU3BJTS1Q8MnMA8op-KWWlOcXZRfzMLCmJeYUp_JCaW4GeTfXEGcPXbBt8QVFmbmJRZXxIFvjwbYaE1YBAN57NlQ</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks</title><source>arXiv.org</source><creator>Dominé, Clémentine C. J ; Anguita, Nicolas ; Proca, Alexandra M ; Braun, Lukas ; Kunin, Daniel ; Mediano, Pedro A. M ; Saxe, Andrew M</creator><creatorcontrib>Dominé, Clémentine C. J ; Anguita, Nicolas ; Proca, Alexandra M ; Braun, Lukas ; Kunin, Daniel ; Mediano, Pedro A. M ; Saxe, Andrew M</creatorcontrib><description>Biological and artificial neural networks develop internal representations that enable them to perform complex tasks. In artificial networks, the effectiveness of these models relies on their ability to build task specific representation, a process influenced by interactions among datasets, architectures, initialization strategies, and optimization algorithms. Prior studies highlight that different initializations can place networks in either a lazy regime, where representations remain static, or a rich/feature learning regime, where representations evolve dynamically. Here, we examine how initialization influences learning dynamics in deep linear neural networks, deriving exact solutions for lambda-balanced initializations-defined by the relative scale of weights across layers. These solutions capture the evolution of representations and the Neural Tangent Kernel across the spectrum from the rich to the lazy regimes. Our findings deepen the theoretical understanding of the impact of weight initialization on learning regimes, with implications for continual learning, reversal learning, and transfer learning, relevant to both neuroscience and practical applications.</description><identifier>DOI: 10.48550/arxiv.2409.14623</identifier><language>eng</language><subject>Computer Science - Learning</subject><creationdate>2024-09</creationdate><rights>http://arxiv.org/licenses/nonexclusive-distrib/1.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,780,885</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2409.14623$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2409.14623$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Dominé, Clémentine C. J</creatorcontrib><creatorcontrib>Anguita, Nicolas</creatorcontrib><creatorcontrib>Proca, Alexandra M</creatorcontrib><creatorcontrib>Braun, Lukas</creatorcontrib><creatorcontrib>Kunin, Daniel</creatorcontrib><creatorcontrib>Mediano, Pedro A. M</creatorcontrib><creatorcontrib>Saxe, Andrew M</creatorcontrib><title>From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks</title><description>Biological and artificial neural networks develop internal representations that enable them to perform complex tasks. In artificial networks, the effectiveness of these models relies on their ability to build task specific representation, a process influenced by interactions among datasets, architectures, initialization strategies, and optimization algorithms. Prior studies highlight that different initializations can place networks in either a lazy regime, where representations remain static, or a rich/feature learning regime, where representations evolve dynamically. Here, we examine how initialization influences learning dynamics in deep linear neural networks, deriving exact solutions for lambda-balanced initializations-defined by the relative scale of weights across layers. These solutions capture the evolution of representations and the Neural Tangent Kernel across the spectrum from the rich to the lazy regimes. Our findings deepen the theoretical understanding of the impact of weight initialization on learning regimes, with implications for continual learning, reversal learning, and transfer learning, relevant to both neuroscience and practical applications.</description><subject>Computer Science - Learning</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNpjYJA0NNAzsTA1NdBPLKrILNMzMjGw1DM0MTMy5mRwcivKz1XwSayqVCjJVwjKTM6wUnCtSEwuUfBJTSzKy8xLV3CpzEvMzUwuVsjMU3BJTS1Q8MnMA8op-KWWlOcXZRfzMLCmJeYUp_JCaW4GeTfXEGcPXbBt8QVFmbmJRZXxIFvjwbYaE1YBAN57NlQ</recordid><startdate>20240922</startdate><enddate>20240922</enddate><creator>Dominé, Clémentine C. J</creator><creator>Anguita, Nicolas</creator><creator>Proca, Alexandra M</creator><creator>Braun, Lukas</creator><creator>Kunin, Daniel</creator><creator>Mediano, Pedro A. M</creator><creator>Saxe, Andrew M</creator><scope>AKY</scope><scope>GOX</scope></search><sort><creationdate>20240922</creationdate><title>From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks</title><author>Dominé, Clémentine C. J ; Anguita, Nicolas ; Proca, Alexandra M ; Braun, Lukas ; Kunin, Daniel ; Mediano, Pedro A. M ; Saxe, Andrew M</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-arxiv_primary_2409_146233</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Computer Science - Learning</topic><toplevel>online_resources</toplevel><creatorcontrib>Dominé, Clémentine C. J</creatorcontrib><creatorcontrib>Anguita, Nicolas</creatorcontrib><creatorcontrib>Proca, Alexandra M</creatorcontrib><creatorcontrib>Braun, Lukas</creatorcontrib><creatorcontrib>Kunin, Daniel</creatorcontrib><creatorcontrib>Mediano, Pedro A. M</creatorcontrib><creatorcontrib>Saxe, Andrew M</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Dominé, Clémentine C. J</au><au>Anguita, Nicolas</au><au>Proca, Alexandra M</au><au>Braun, Lukas</au><au>Kunin, Daniel</au><au>Mediano, Pedro A. M</au><au>Saxe, Andrew M</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks</atitle><date>2024-09-22</date><risdate>2024</risdate><abstract>Biological and artificial neural networks develop internal representations that enable them to perform complex tasks. In artificial networks, the effectiveness of these models relies on their ability to build task specific representation, a process influenced by interactions among datasets, architectures, initialization strategies, and optimization algorithms. Prior studies highlight that different initializations can place networks in either a lazy regime, where representations remain static, or a rich/feature learning regime, where representations evolve dynamically. Here, we examine how initialization influences learning dynamics in deep linear neural networks, deriving exact solutions for lambda-balanced initializations-defined by the relative scale of weights across layers. These solutions capture the evolution of representations and the Neural Tangent Kernel across the spectrum from the rich to the lazy regimes. Our findings deepen the theoretical understanding of the impact of weight initialization on learning regimes, with implications for continual learning, reversal learning, and transfer learning, relevant to both neuroscience and practical applications.</abstract><doi>10.48550/arxiv.2409.14623</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2409.14623
ispartof
issn
language eng
recordid cdi_arxiv_primary_2409_14623
source arXiv.org
subjects Computer Science - Learning
title From Lazy to Rich: Exact Learning Dynamics in Deep Linear Networks
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-09T23%3A04%3A06IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=From%20Lazy%20to%20Rich:%20Exact%20Learning%20Dynamics%20in%20Deep%20Linear%20Networks&rft.au=Domin%C3%A9,%20Cl%C3%A9mentine%20C.%20J&rft.date=2024-09-22&rft_id=info:doi/10.48550/arxiv.2409.14623&rft_dat=%3Carxiv_GOX%3E2409_14623%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true