Scalable kernel methods using randomized numerical linear algebra

dc.contributor.advisorGonzalez Osorio, Fabio Augusto
dc.contributor.authorCastellanos Martinez, Ivan Yesid
dc.contributor.researchgroupMindLabspa
dc.date.accessioned2021-11-18T04:30:09Z
dc.date.available2021-11-18T04:30:09Z
dc.date.issued2021
dc.descriptionDocumento de tesis de maestriaspa
dc.descriptionilustraciones, tablasspa
dc.description.abstractLos métodos de kernel corresponden a un grupo de algoritmos de aprendizaje maquinal que hacen uso de una función de kernel para representar implicitamente datos en un espacio de alta dimensionalidad, donde sistemas de optimización lineal guíen a relaciones no lineales en el espacio original de los datos y por lo tanto encontrando patrones complejos dento de los datos. La mayor desventaja que tienen estos métodos es su pobre capacidad de escalamiento, pues muchos algoritmos basados en kernel requiren calcular una matriz de orden cuadrática respecto al numero de ejemplos en los datos, esta limitación ha provocado que los metodos de kernel sean evitados en configuraciones de datos a gran escala y utilicen en su lugar tecnicas como el aprendizaje profundo. Sin embargo, los metodos de kernel todavía son relevantes para entender mejor los métodos de aprendizaje profundo y ademas pueden mejorarlos haciendo uso de estrategias híbridas que combinen lo mejor de ambos mundos. El principal objetivo de esta tesis es explorar maneras eficientes de utilizar métodos de kernel sin una gran pérdida en precisión. Para realizar esto, diferentes enfoque son presentados y formulados, dentro de los cuales, nosotros proponemos la estrategía de aprendizaje utilizando budget, la cual es presentada en detalle desde una perspectiva teórica, incluyendo un procedimiento novedoso para la selección del budget, esta estrategia muestra en la evaluación experimental un rendimiento competitivo y mejoras respecto al método estandar de aprendizaje utilizando budget, especialmente cuando se seleccionan aproximaciones mas pequeñas, las cuales son las mas útiles en ambientes de gran escala. (Texto tomado de la fuente)spa
dc.description.abstractKernel methods are a set of machine learning algorithms that make use of a kernel function in order to represent data in an implicit high dimensional space, where linear optimization systems lead to non-linear relationships in the data original space and therefore finding complex patterns in the data. The main disadvantage of these methods is their poor scalability, as most kernel based algorithms need to calculate a matrix of quadratic order regarding the number of data samples. This limitation has caused kernel methods to be avoided for large scale datasets and use approaches such as deep learning instead. However, kernel methods are still relevant to better understand deep learning methods and can improve them through hybrid settings that combine the best of both worlds. The main goal of this thesis is to explore efficient ways to use kernel methods without a big loss in accuracy performance. In order to do this, different approaches are presented and formulated, from which, we propose the learning-on-a-budget strategy, which is presented in detail from a theoretical perspective, including a novel procedure of budget selection. This strategy shows, in the experimental evaluation competitive performance and improvements to the standard learning-on-a-budget method, especially when selecting smaller approximations, which are the most useful in large scale environments.eng
dc.description.degreelevelMaestríaspa
dc.description.degreenameMagíster en Ingeniería - Ingeniería de Sistemas y Computaciónspa
dc.description.researchareaCiencias de la computaciónspa
dc.format.extentxv, 58 páginasspa
dc.format.mimetypeapplication/pdfspa
dc.identifier.instnameUniversidad Nacional de Colombiaspa
dc.identifier.reponameRepositorio Institucional Universidad Nacional de Colombiaspa
dc.identifier.repourlhttps://repositorio.unal.edu.co/spa
dc.identifier.urihttps://repositorio.unal.edu.co/handle/unal/80695
dc.language.isoengspa
dc.publisherUniversidad Nacional de Colombiaspa
dc.publisher.branchUniversidad Nacional de Colombia - Sede Bogotáspa
dc.publisher.departmentDepartamento de Ingeniería de Sistemas e Industrialspa
dc.publisher.facultyFacultad de Ingenieríaspa
dc.publisher.placeBogotá, Colombiaspa
dc.publisher.programBogotá - Ingeniería - Maestría en Ingeniería - Ingeniería de Sistemas y Computaciónspa
dc.relation.referencesAhuja, S. and Angra, S. (2017). Machine learning and its applications: A review.spa
dc.relation.referencesBaveye, Y., Dellandr´ea, E., Chamaret, C., and Chen, L. (2015). Deep learning vs. kernel methods: Performance for emotion prediction in videos. In 2015 International Conference on Affective Computing and Intelligent Interaction (ACII), pages 77–83.spa
dc.relation.referencesBelkin, M., Ma, S., and Mandal, S. (2018). To understand deep learning we need to understand kernel learning.spa
dc.relation.referencesBengio, Y., Delalleau, O., and Le Roux, N. (2005). The curse of dimensionality for local kernel machines. Techn. Rep, 1258:12.spa
dc.relation.referencesBorgwardt, K. M. (2011). Kernel Methods in Bioinformatics, pages 317– 334. Springer Berlin Heidelberg, Berlin, Heidelberg.spa
dc.relation.referencesBousquet, O. and Herrmann, D. J. (2003). On the complexity of learning the kernel matrix. Advances in neural information processing systems, pages 415–422.spa
dc.relation.referencesBoutsidis, C., Mahoney, M. W., and Drineas, P. (2009). An Improved Approximation Algorithm for the Column Subset Selection Problem, pages 968–977.spa
dc.relation.referencesChen, D., Jacob, L., and Mairal, J. (2020). Convolutional kernel networks for graph-structured data. In III, H. D. and Singh, A., editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 1576–1586. PMLR.spa
dc.relation.referencesChitta, R., Jin, R., Havens, T. C., and Jain, A. K. (2014). Scalable kernel clustering: Approximate kernel k-means. CoRR, abs/1402.3849.spa
dc.relation.referencesChitta, R., Jin, R., and Jain, A. K. (2012). Efficient kernel clustering using random fourier features. In 2012 IEEE 12th International Conference on Data Mining, pages 161–170. IEEE.spa
dc.relation.referencesWu, L., Chen, P.-Y., Yen, I. E.-H., Xu, F., Xia, Y., and Aggarwal, C. (2018). Scalable spectral clustering using random binning features. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pages 2506–2515.spa
dc.relation.referencesXiao, H., Rasul, K., and Vollgraf, R. (2017). Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms.spa
dc.relation.referencesYang, M.-H. (2002). Kernel eigenfaces vs. kernel fisherfaces: Face recognition using kernel methods. In Fgr, volume 2, page 215.spa
dc.relation.referencesYu, F. X., Suresh, A. T., Choromanski, K., Holtmann-Rice, D. N., and Kumar, S. (2016). Orthogonal random features. CoRR, abs/1610.09072.spa
dc.relation.referencesZhang, D., Wang, J., Cai, D., and Lu, J. (2010). Self-taught hashing for fast similarity search. In Proceedings of the 33rd international ACM SIGIR conference on Research and development in information retrieval, pages 18–25.spa
dc.relation.referencesZhang, Q., Filippi, S., Gretton, A., and Sejdinovic, D. (2017). Largescale kernel methods for independence testing. Statistics and Computing, 28(1):113–130.spa
dc.relation.referencesTrokicic, A. and Todorovic, B. (2020). On expected error of randomized nystrom kernel regression. Filomat, 34(11):3871–3884.spa
dc.relation.referencesVanegas, J. A., Escalante, H. J., and González, F. A. (2018). Semisupervised online kernel semantic embedding for multi-label annotation. In Mendoza, M. and Velastín, S., editors, Progress in Pattern Recognition, Image Analysis, Computer Vision, and Applications, pages 693–701, Cham. Springer International Publishing.spa
dc.relation.referencesWang, D. E. J. (2006). Fast approximation of centrality. Graph algorithms and applications, 5(5):39.spa
dc.relation.referencesWang, J., Cao, B., Yu, P., Sun, L., Bao, W., and Zhu, X. (2018). Deep learning towards mobile applications. In 2018 IEEE 38th International Conference on Distributed Computing Systems (ICDCS), pages 1385–1393.spa
dc.relation.referencesWang, J., Shen, H. T., Song, J., and Ji, J. (2014). Hashing for similarity search: A survey. CoRR, abs/1408.2927.spa
dc.relation.referencesWang, S., Gittens, A., and Mahoney, M. W. (2019). Scalable kernel k-means clustering with nyström approximation: relative-error bounds. The Journal of Machine Learning Research, 20(1):431–479.spa
dc.relation.referencesWang, S., Luo, L., and Zhang, Z. (2016). Spsd matrix approximation via column selection: Theories, algorithms, and extensions.spa
dc.relation.referencesWang, S. and Zhang, Z. (2013). Improving cur matrix decomposition and the nyström approximation via adaptive sampling. The Journal of Machine Learning Research, 14(1):2729–2769.spa
dc.relation.referencesWang, Y., Liu, X., Dou, Y., Lv, Q., and Lu, Y. (2017). Multiple kernel learning with hybrid kernel alignment maximization. Pattern Recognition, 70:104–111.spa
dc.relation.referencesWang, Z., Crammer, K., and Vucetic, S. (2012). Breaking the curse of kernelization: Budgeted stochastic gradient descent for large-scale svm training. Journal of Machine Learning Research, 13(100):3103–3131.spa
dc.relation.referencesWilliams, C. K. I. and Seeger, M. (2001). Using the nyström method to speed up kernel machines. In Leen, T. K., Dietterich, T. G., and Tresp, V., editors, Advances in Neural Information Processing Systems 13, pages 682–688. MIT Press.spa
dc.relation.referencesWitten, R. and Candes, E. (2015). Randomized algorithms for low-rank matrix factorizations: sharp performance bounds. Algorithmica, 72(1):264–281.spa
dc.rights.accessrightsinfo:eu-repo/semantics/openAccessspa
dc.rights.licenseReconocimiento 4.0 Internacionalspa
dc.rights.urihttp://creativecommons.org/licenses/by/4.0/spa
dc.subject.ddc000 - Ciencias de la computación, información y obras generalesspa
dc.subject.proposalMachine Learningeng
dc.subject.proposalKernel Methodseng
dc.subject.proposalBudget Methodeng
dc.subject.proposalRandomized Numerical Linear Algebraeng
dc.subject.proposalDistance Based Hashingeng
dc.subject.proposalApproximated Methodseng
dc.subject.proposalAprendizaje maquinalspa
dc.subject.proposalMétodos de kernelspa
dc.subject.proposalMétodo de budgetspa
dc.subject.proposalAlgebra Lineal Numérica Aleatorizadaspa
dc.subject.proposalHashing basado en distanciasspa
dc.subject.proposalMétodos Aproximadosspa
dc.titleScalable kernel methods using randomized numerical linear algebraeng
dc.title.translatedMétodos de kernel escalables utilizando álgebra lineal numérica aleatorizadaspa
dc.typeTrabajo de grado - Maestríaspa
dc.type.coarhttp://purl.org/coar/resource_type/c_bdccspa
dc.type.coarversionhttp://purl.org/coar/version/c_ab4af688f83e57aaspa
dc.type.contentTextspa
dc.type.driverinfo:eu-repo/semantics/masterThesisspa
dc.type.redcolhttp://purl.org/redcol/resource_type/TMspa
dc.type.versioninfo:eu-repo/semantics/acceptedVersionspa
dcterms.audience.professionaldevelopmentPúblico generalspa
oaire.accessrightshttp://purl.org/coar/access_right/c_abf2spa

Archivos

Bloque original

Mostrando 1 - 1 de 1
Cargando...
Miniatura
Nombre:
1032463787.2021.pdf
Tamaño:
1019.48 KB
Formato:
Adobe Portable Document Format
Descripción:
Tesis de Maestría en Ingeniería de Sistemas y Computación

Bloque de licencias

Mostrando 1 - 1 de 1
Cargando...
Miniatura
Nombre:
license.txt
Tamaño:
3.98 KB
Formato:
Item-specific license agreed upon to submission
Descripción: