Biologically Plausible Learning with Memory
Biologicky vÄ›rohodnĂ© uÄŤenĂ s pamÄ›tĂ
Authors
Supervisors
Reviewers
Editors
Other contributors
Journal Title
Journal ISSN
Volume Title
Publisher
České vysoké učenà technické v Praze
Czech Technical University in Prague
Czech Technical University in Prague
Date of defense
2025-06-19
Abstract
Online Continual learning (OCL) je pro-
blĂ©m, ve kterĂ©m modely strojovĂ©ho uÄŤenĂ
fungujà dobře jen pokud se umà dobře učit
postupně z proudu dat, která nejsou ná-
hodnÄ› zamĂchaná a rovnomÄ›rnÄ› rozdÄ›lená.
OCL to dÄ›lá rozdÄ›lenĂm datovĂ© sady na
několik úkolů, z nichž každý má nějakou
část tĹ™Ăd. Tyto Ăşkoly se pak dávajĂ mo-
delu jeden po druhĂ©m, pĹ™edstĂrajĂcĂ proud
pĹ™ĂkladĹŻ, konÄŤĂcĂ s datasetem.
NejvÄ›tšĂm problĂ©mem v OCL je cata-
strophic forgetting(katastrofálnĂ zapomĂ-
nánĂ). Ten se vykazuje tĂm, Ĺľe model,
který je naučen běžným způsobem, si ne-
může dobře pamatovat předchozà úkoly a
nedokáže je pak dobře řešit.
Tahle práce k tomu přidává tři pod-
mĂnky, kterĂ© normálnÄ› nejsou povinnĂ©: (i)
velikost batchů je jedna, (ii) omezená veli-
kost paměti, a (iii) žádné informace o tom,
kdy se úkoly měnà nebo jaký je právě ak-
tivnĂ Ăşkol.
Ve studii se řešà tři metody GSS, ER a
DER++. Jsou porovány podle toho, jak
si pamatujĂ data, jak promĂchávajĂ stream
dat a jak se snažà zabránit zapomĂnanĂ.
Výsledkem experimentů je poznatek,
že žádná z metod to nedělá úplně dobře.
Všechny majĂ problĂ©m správnÄ› namĂchat
zapamatovaná data tak, aby stream při-
pomĂnal náhodnĂ© rozdÄ›lenĂ.
Na zlepšenà těchto problémů jsou navr-
Ĺľeny tĹ™i nová Ĺ™ešenĂ: zaprvĂ© metoda na mĂ-
chánà zapamatovaných exemplářů do stre-
amu dat, zadruhĂ© strategie na vytvářenĂ
ĂşkolĹŻ, kterĂ© jsou vĂce závislĂ©, a nakonec
nápad jak vyuĹľĂt pĹ™edtrĂ©novanĂ˝ model,
aby to celĂ© fungovalo lĂp.
Online Continual Learning (OCL) intro- duces a setting where machine learning models succeed only when they can learn sequentially from a non-independent and identically distributed (i.i.d) stream of data. OCL ensures this by splitting a dataset into different tasks, each containing a sub- set of the classes, and feeding them to the model as a continuous stream until the entire dataset is processed. The most prominent problem in OCL is catastrophic forgettingthe models fail- ure to retain performance on earlier tasks. This thesis adds three challenging con- straints to OCL: (i) batch size of one, (ii) finite memory size, and (iii) a task-free policy where models dont know when tasks change or what the current task is. Through an empirical study of the methods GSS, ER and DER++ a com- parison is made of how they manage their memory, how they transform the stream of data and how they they try to mitigate catastrophic forgetting. The study concludes, that neither of the methods are completely successful in solving the problem that has been set, as they all fail in the way of mixing the remembered data correctly to create an i.i.d. stream. To address this, three novel solutions are proposed: a method for mixing memo- rized samples, a strategy for creating more dependent tasks, and an inquiry into using pre-trained models for improved perfor- mance.
Online Continual Learning (OCL) intro- duces a setting where machine learning models succeed only when they can learn sequentially from a non-independent and identically distributed (i.i.d) stream of data. OCL ensures this by splitting a dataset into different tasks, each containing a sub- set of the classes, and feeding them to the model as a continuous stream until the entire dataset is processed. The most prominent problem in OCL is catastrophic forgettingthe models fail- ure to retain performance on earlier tasks. This thesis adds three challenging con- straints to OCL: (i) batch size of one, (ii) finite memory size, and (iii) a task-free policy where models dont know when tasks change or what the current task is. Through an empirical study of the methods GSS, ER and DER++ a com- parison is made of how they manage their memory, how they transform the stream of data and how they they try to mitigate catastrophic forgetting. The study concludes, that neither of the methods are completely successful in solving the problem that has been set, as they all fail in the way of mixing the remembered data correctly to create an i.i.d. stream. To address this, three novel solutions are proposed: a method for mixing memo- rized samples, a strategy for creating more dependent tasks, and an inquiry into using pre-trained models for improved perfor- mance.
Description
Citation
Underlying research data set URL
Permanent link
Rights/License
Vysokoškolská závÄ›reÄŤná práce je dĂlo chránÄ›nĂ© autorskĂ˝m zákonem. Je moĹľnĂ© poĹ™izovat z nÄ›j na svĂ© náklady a pro svoji osobnĂ potĹ™ebu vĂ˝pisy, opisy a rozmnoĹľeniny. Jeho vyuĹľitĂ musĂ bĂ˝t v souladu s autorskĂ˝m zákonem v platnĂ©m znÄ›nĂ.
A university thesis is a work protected by the Copyright Act of the Czech Republic. Extracts, copies and transcripts of the thesis are allowed for personal use only and at one`s own expense. The use of thesis should be in compliance with the Copyright Act.
A university thesis is a work protected by the Copyright Act of the Czech Republic. Extracts, copies and transcripts of the thesis are allowed for personal use only and at one`s own expense. The use of thesis should be in compliance with the Copyright Act.