On the optimization of restricted Boltzmann machines and its application for fingerprint-based biometrics

• Grant number: 14/24491-6
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: December 20, 2014
• End date: January 30, 2015
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Aparecido Nilceu Marana
• Grantee: Gustavo Henrique de Rosa
• Supervisor: David Cox
• Host Institution: Faculdade de Ciências (FC). Universidade Estadual Paulista (UNESP). Campus de Bauru. Bauru , SP, Brazil
• Institution abroad: Harvard University, Cambridge, United States
• Associated to the scholarship: 14/09125-3 - Convolutional neural networks applied to the biometrics person recognition, BP.IC

Abstract

Deep learning techniques have been widely used in the last years due to their promising results in several applications, mainly face and object detection. However, one of their main shortcomings is related to the selection of suitable parameters in order to allow reasonable results. Since we may have millions of parameters, a manual fine tuning of them seems to be impractical. In this research proposal, we deal with this problem by means of meta-heuristic-based techniques, specifically Harmony Search, since it has been consistently efficient in several other applications studied by the research group to which the student is related to. Despite we can face several deep learning-based techniques out there, we opted to study Restricted Boltzmann Machines due to their high ability in learning good data representations. Additionally, the proposed work is validated in the context of fingerprint-based person recognition, as stated in the student's main project. (AU)