Speaker Biography
Haroldo Fraga de Campos Velho received a B. S. in Chemical Engineering from the Pontificia Universidade Catolica do Rio Grande do Sul (PUCRS), Brazil, in 1982, D.Sc. (emphasis on computational fluid dynamics) and M.Sc. degrees (emphasis on nuclear engineering: transport theory) in Mechanical Engineering from the Federal University of the Rio Grande do Sul (UFRGS), Brazil, in 1992 and 1988, respectively. He was a visiting scientist for the Institute of Cosmo-Geophysics (Turin, Italy) – 1997, and Department of Atmospheric Science from the Colorado State University (CO, USA) – 1998. He was elected as a Member of the Council of the Brazilian Society of Computational and Applied Mathematics (2002-2006). He was the Associate Director for Space and Environment of the INPE for the period: 2008-2010. Currently, he is a senior researcher of the Associate Laboratory for Computing and Applied Mathematics (LABAC) of the National Institute for Space Research (INPE). His research activity is in scientific computing, development and applications of inverse problems in several topics of space research. Other research activities are in atmospheric turbulence modelling, data assimilation using artificial neural networks and adaptive particle filter.
Talk Abstract
Machine learning is one of key issues for the modern data science framework. Among techniques of machine learning, neural networks has emerged to be used for several applications. In the seminar, a self-configured supervised neural network is presented. The best architecture to the neural network is formulated as an optimization problem. Basically, the user provides the data, and the system replies with a neural network with optimal architecture for a given application. Two examples will be presented in the seminar: autonomous navigation to the Unmanned Aerial Vehicle (UAV), applying image processing for estimating the drone position neural network implemented on FPGA.; and data assimilation, where a set of neural networks were trained to emulate local ensemble transform Kalman filter for weather prediction models