Development of Support Vector Machines (SVMs) in Graphics Processing Units for Pattern Recognition
resumo_portugues_corrigido.pdf
The Master of Science thesis where I developed a Support Vector Machine (SVM) which uses the power of modern GPUs. Therefore, the learning and classification tasks which could take weeks are reduced to hours or even less, depending on the power of the GPU. Not only it allows bigger learning tasks to be handled but it allows fine tuning of the kernel parameters and penalization constant, a process known as grid search which is computationally expensive.
It was written in C++ and requires NVIDIA's CUDA, as well as a NVIDIA GPU with unified shaders (GeForce 8 or above). The SVM is included in the open-source machine learning library GPUMLib managed by the advisors [Bernardete Ribeiro] and [Noel Lopes]. The library can be freely downloaded from [sourceforge].
The classifier was validated in famous datasets found on the UCI machine learning [repository] against the state-of-the-art LIBSVM and found to be equivalent in classification performance, while being faster.
The second part of my thesis comprised of using the previously developed SVM-GPU classifier to study three real world problems: protein prediction, MP3 Steganalysis and the main focus of the chapter, the offline recognition of handwritten signatures. This problem consisted of three sub-problems: the recognition of signature author, a generic detection of forged/original signatures and the detection if for a given person, the scanned signature is authentic or forged. We achieved excellent results in the identification of the signature's author (F-Score of 98.14%). For the forged/original signature detection per author we obtained an average F-Score of 71.62% which is interesting, but implicates further research in this field of forgery detection.
Publications in International Conferences
João Gonçalves, Noel Lopes and Bernardete Ribeiro, "Multi-Threaded Support Vector Machines for Pattern Recognition", in International Conference on Neural Information Processing, LNCS, Springer, November 2012, 2012
Noel Lopes, Bernardete Ribeiro and João Gonçalves, "Restricted Boltzman Machines and Deep Belief Networks on Multi-Core Processors", in IEEE World Congress on Computational Intelligence (WCCI 2012), IEEE, Brisbane, Australia, 2012