Sparse and Non-Sparse Multiple Kernel Learning for Recognition
Aprendizaje de múltiples núcleos esparcidos y no esparcidos para reconocimiento
Mitchel Alioscha-Pérez 1, Hichem Sahli 2, Isabel González 2, and Alberto Taboada-Crispi 1
1 Centro de Estudios de Electrónica y Tecnologias de la Información (CEETI), Universidad Central de Las Villas, Villa Clara, Cuba sirmichel@gmail.com, ataboada@uclv.edu.cu
]]> 2 Department of Electronics and Informatics (ETRO), Vrije Universiteit Brussel, Brussels, Belgium hichem.sahl@etro.vub.ac.bei, igonzale@etro.vub.ac.be
Article received on 02/02/2011.
Accepted on 05/10/2011.
Abstract
The development of Multiple Kernel Techniques has become of particular interest for machine learning researchers in Computer Vision topics like image processing, object classification, and object state recognition. Sparsity-inducing norms along with non-sparse formulations promote different degrees of sparsity at the kernel coefficient level, at the same time permitting non-sparse combination within each individual kernel. This makes MKL models very suitable for different problems, allowing adequate selection of the regularizer according to different norms and the nature of the problem. We formulate and discuss MKL regularizations and optimization approaches, as well as demonstrate MKL effectiveness compared to the state-of-the-art SVM models using a Computer Vision Recognition problem.
Keywords. Multiple kernel learning, object state recognition, norm regularizers, analytical updates, cutting plane method, Newton's method.
Resumen
]]> El desarrollo de técnicas MKL (aprendizaje de múltiples núcleos) ha sido de particular interés para los investigadores en el aprendizaje automatizado, en tópicos de visión por computadora, así como para el procesamiento de imágenes, clasificación de objetos, y reconocimiento del estado de los objetos. En ecuaciones donde se combinan múltiples núcleos, las normas de inducción de dispersión junto con las formulaciones de no dispersión, promueven diferentes grados de dispersión a nivel de los coeficientes de combinación, mientras que permiten la combinación no esparcida en los núcleos individuales. Esto hace de los modelos MKL muy adecuados para diferentes problemas, permitiendo la selección óptima del regularizador, y así lograr un mejor reconocimiento de acuerdo a la naturaleza del problema. En este trabajo, formulamos y discutimos las diferentes regularizaciones de MKL y los métodos de optimización relacionados, demostrando su efectividad en un problema de reconocimiento de visión por computadora.Palabras clave. Aprendizaje de múltiples núcleos, reconocimiento del estado de objetos, regularizadores de normas, actualizaciones analíticas, método de planos cortantes, método de Newton.
DESCARGAR ARTÍCULO EN FORMATO PDF
References
1. Bach, F., Lanckriet, G.R.G., & Jordan, M.I. (2004). Multiple Kernel Learning, Conic Duality, and the SMO Algorithm. Twenty-first international conference on Machine learning (ICML '04), Banff, Alberta, Canada. [ Links ]
2. Chapelle, O. & Rakotomamonjy, A. (2008). Second order optimization of kernel parameters. NIPS´08 Workshop: Kernel Learning - Automatic Selection of Optimal Kernels, Whistler, Canada. [ Links ]
]]>3. Ekman, P., Friesen, W.V., & Hager, J.C. (2002). The Facial Action Coding System: The Manual on CD-ROM & Investigator's Guide (2nd Ed.). Weston, Florida: Network Information Research Corporation. [ Links ]
4. Gonzalez, I., Sahli, H., & Verhelst, W. (2010). Automatic Recognition of Lower Facial Action Units. 7th International Conference on Methods and Techniques in Behavioral Research (MB'10), Eindhoven, Netherlands, Article No.8. [ Links ]
5. Hou, Y., Sahli, H., Ilse, R., Zhang, Y., & Zhao, R. (2007). Robust Shape-based Head Tracking. Advanced Concepts for Intelligent Vision Systems (ACIVS 2007), Lecture Notes in Computer Science, 4678, 340–351. [ Links ]
6. Kloft, M., Brefeld, U., Laskov, P., & Sonnenburg, S. (2008). Non-sparse multiple kernel learning. NIPS'08 Workshop: Kernel Learning - Automatic Selection of Optimal Kernels, Whistler, Canada. [ Links ]
7. Kloft, M., Brefeld, U., Sonnenburg, S., & Zien, A. (2010). Non-sparse regularization and efficient training with multiple kernels (Technical Report No. UCB/EECS-2010-21). Berkeley: University of California at Berkeley. [ Links ]
]]>8. Kloft, M., Brefeld, U., Sonnenburg, S., Laskov, P., Müller K.R., & Zien, A. (2009). Efficient and accurate Lp-norm multiple kernel learning. Advances in Neural Information Processing Systems, 22, 997–1005. [ Links ]
9. Lanckriet, G.R.G., Cristianini, N., Bartlett, P., Ghaoui, L.E., & Jordan, M.I. (2004). Learning the Kernel Matrix with Semidefinite Programming. The Journal of Machine Learning Research, 5(Jan), 27–72. [ Links ]
10. Rakotomamonjy, A., Bach, F.R., Canu, S., & Grandvalet, Y. (2008). SimpleMKL. Journal of Machine Learning Research, 9(Nov), 2491–2521. [ Links ]
11. Schölkopf, B. & Smola, A.J. (2002). Learning with Kernels: Support Vector Machines, Regularization, Optimization, and Beyond. Cambridge, Mass.: MIT Press. [ Links ]
12. Sonnenburg, S., Rätsch, G., Schäfer, C., & Schölkopf, B. (2006). Large Scale Multiple Kernel Learning. Journal of Machine Learning Research, 7(Jul), 1531–1565. [ Links ]
]]>13. Xu, Z., Jin, R., Yang, H., King, I., & Lyu, M.R. (2010). Simple and Efficient Multiple Kernel Learning by Group Lasso. 27th International Conference on Machine Learning (ICML 2010), Haifa, Israel, Paper ID: 540. [ Links ]
]]>