Active learning of regular languages as an approach to neural language models verification.

This work tackles the general problem of verifying the behavior of sequence processing neural networks, specifically neural acceptors and neural language models. The contribution is a framework for extracting formal abstractions of the networks under analysis and verifying whether they satisfy given...

Celý popis

Uloženo v:
Podrobná bibliografie
Hlavní autor: Mayr, Franz (author)
Médium: doctoralThesis
Jazyk:angličtina
Vydáno: 2024
On-line přístup:https://hdl.handle.net/20.500.12008/45940
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
_version_ 1868890225177001984
author Mayr, Franz
author_browse Mayr, Franz
author_facet Mayr, Franz
author_role author
collection COLIBRI
dc.contributor.none.fl_str_mv Mayr Franz, Universidad de la República (Uruguay). Facultad de Ingeniería.
dc.creator.none.fl_str_mv Mayr, Franz
dc.date.none.fl_str_mv 2024-09-18T15:58:38Z
2024-09-18T15:58:38Z
2024
dc.format.none.fl_str_mv 105 p.
application/pdf
dc.identifier.none.fl_str_mv Mayr, F. Active learning of regular languages as an approach to neural language models verification [en línea] Tesis de doctorado. Udelar. FI. INCO. PEDECIBA. Área Informática, 2024.
https://hdl.handle.net/20.500.12008/45940
dc.language.none.fl_str_mv en
eng
dc.publisher.none.fl_str_mv Udelar. FI.
dc.rights.none.fl_str_mv info:eu-repo/semantics/openAccess
Licencia Creative Commons Atribución - No Comercial - Sin Derivadas (CC - By-NC-ND 4.0)
dc.source.none.fl_str_mv reponame:COLIBRI
instname:Universidad de la República
instacron:Universidad de la República
dc.title.none.fl_str_mv Active learning of regular languages as an approach to neural language models verification.
dc.type.none.fl_str_mv Tesis de doctorado
info:eu-repo/semantics/doctoralThesis
info:eu-repo/semantics/acceptedVersion
description This work tackles the general problem of verifying the behavior of sequence processing neural networks, specifically neural acceptors and neural language models. The contribution is a framework for extracting formal abstractions of the networks under analysis and verifying whether they satisfy given requirements. This process steps on two core ideas: 1) treating the neural network as a black box, and 2) using a probabilistic framework to analyze how much the extracted model approximates the original one. For this matter, a series of active learning algorithms and techniques are proposed, developed and analyzed. For the case of neural acceptors a procedure for checking properties of neural networks is presented. This approach is able to check properties without explicitly building representations of the network. We show that this approach offers better guarantees and is more efficient than post-learning verification where the property is checked on a learned model of the network. Besides, it does not require resorting to an external decision procedure for verification nor fixing a specific requirement specification formalism. For the case of neural language models, a learning algorithm based on a congruence over strings which is parameterized by an equivalence relation over probability distributions is presented. The learning algorithm is implemented using a tree data structure and shown to be empirically more efficient than reference techniques.
eu_rights_str_mv openAccess
format doctoralThesis
id anni_ecaf9f82228688eaa498231d3feb0a64
identifier_str_mv Mayr, F. Active learning of regular languages as an approach to neural language models verification [en línea] Tesis de doctorado. Udelar. FI. INCO. PEDECIBA. Área Informática, 2024.
instacron_str Universidad de la República
institution Universidad de la República
instname_str Universidad de la República
language eng
language_invalid_str_mv en
network_acronym_str anni
network_name_str oai-lr-anni
oai_identifier_str oai:colibri.udelar.edu.uy:20.500.12008/45940
publishDate 2024
publishDateSort 2024
publisher.none.fl_str_mv Udelar. FI.
reponame_str COLIBRI
repository.mail.fl_str_mv
repository.name.fl_str_mv
repository_id_str
rights_invalid_str_mv Licencia Creative Commons Atribución - No Comercial - Sin Derivadas (CC - By-NC-ND 4.0)
spelling Active learning of regular languages as an approach to neural language models verification.Mayr, FranzThis work tackles the general problem of verifying the behavior of sequence processing neural networks, specifically neural acceptors and neural language models. The contribution is a framework for extracting formal abstractions of the networks under analysis and verifying whether they satisfy given requirements. This process steps on two core ideas: 1) treating the neural network as a black box, and 2) using a probabilistic framework to analyze how much the extracted model approximates the original one. For this matter, a series of active learning algorithms and techniques are proposed, developed and analyzed. For the case of neural acceptors a procedure for checking properties of neural networks is presented. This approach is able to check properties without explicitly building representations of the network. We show that this approach offers better guarantees and is more efficient than post-learning verification where the property is checked on a learned model of the network. Besides, it does not require resorting to an external decision procedure for verification nor fixing a specific requirement specification formalism. For the case of neural language models, a learning algorithm based on a congruence over strings which is parameterized by an equivalence relation over probability distributions is presented. The learning algorithm is implemented using a tree data structure and shown to be empirically more efficient than reference techniques.El presente trabajo aborda el problema general de la verificación del comportamiento de redes neuronales que procesan secuencias, en concreto los aceptores neuronales y los modelos neuronales de lenguaje. La tesis desarrolla un marco teórico-práctico para la extracción de abstracciones formales y la verificación de las redes neuronales bajo análisis. Este proceso se basa en dos ideas centrales: 1) tratar la red neuronal como una caja negra, y 2) utilizar un marco probabilístico para analizar en que medida el modelo extraído se aproxima al original. Para ello, se proponen, desarrollan y analizan una serie de algoritmos y técnicas de aprendizaje activo. Para el caso de los aceptores neuronales se presenta un procedimiento de verificación de propiedades de redes neuronales. Este enfoque es capaz de verificar propiedades sin construir explícitamente representaciones de la red. Se demuestra que este enfoque ofrece mejores garantías y es mas eficiente que la verificación posterior al aprendizaje, en la que la propiedad se verifica únicamente sobre el modelo aprendido de la red. Además, no requiere recurrir a un procedimiento de decisión externo para la verificación ni fijar un formalismo específico de especificación de requisitos. Para el caso de los modelos neuronales de lenguaje se presenta un algoritmo de aprendizaje basado en una congruencia sobre secuencias que se parametriza mediante una relación de equivalencia sobre distribuciones de probabilidad. El algoritmo de aprendizaje se implementa utilizando una estructura de datos en árbol y se muestra que es empíricamente mas eficiente que las técnicas de referencia.Udelar. FI.Mayr Franz, Universidad de la República (Uruguay). Facultad de Ingeniería.2024-09-18T15:58:38Z2024-09-18T15:58:38Z2024Tesis de doctoradoinfo:eu-repo/semantics/doctoralThesisinfo:eu-repo/semantics/acceptedVersion105 p.application/pdfMayr, F. Active learning of regular languages as an approach to neural language models verification [en línea] Tesis de doctorado. Udelar. FI. INCO. PEDECIBA. Área Informática, 2024.https://hdl.handle.net/20.500.12008/45940reponame:COLIBRIinstname:Universidad de la Repúblicainstacron:Universidad de la RepúblicaenengLas obras depositadas en el Repositorio se rigen por la Ordenanza de los Derechos de la Propiedad Intelectual de la Universidad de la República.(Res. Nº 91 de C.D.C. de 8/III/1994 – D.O. 7/IV/1994) y por la Ordenanza del Repositorio Abierto de la Universidad de la República (Res. Nº 16 de C.D.C. de 07/10/2014)info:eu-repo/semantics/openAccessLicencia Creative Commons Atribución - No Comercial - Sin Derivadas (CC - By-NC-ND 4.0)oai:colibri.udelar.edu.uy:20.500.12008/459402026-04-14T10:27:58Z
spellingShingle Active learning of regular languages as an approach to neural language models verification.
Mayr, Franz
status_str acceptedVersion
title Active learning of regular languages as an approach to neural language models verification.
title_full Active learning of regular languages as an approach to neural language models verification.
title_fullStr Active learning of regular languages as an approach to neural language models verification.
title_full_unstemmed Active learning of regular languages as an approach to neural language models verification.
title_short Active learning of regular languages as an approach to neural language models verification.
title_sort Active learning of regular languages as an approach to neural language models verification.
url https://hdl.handle.net/20.500.12008/45940