Juan-Pablo Carbajal - Data Science
Date: - Location: Eurecom
Abstract: Machine learning techniques have been developed and optimized to model data generating process of unknown nature for which we have almost no insight about their mechanisms. It offers powerful and automatic modeling methods, but generally data-hungry. Scientific modeling lays on the other extreme, we build mechanistic models to describe sparse data generated by experiments and observations. The latter is a fruitful but expensive process, which is generally inapposite for big data or complex systems scenarios. Is it possible to apply scientific modeling to large datasets? Can we exploit automatic modeling methods in sparse data scenarios? The intersection between machine learning and scientific modeling promises an answer to these questions. Although this has been identified at least since the beginning of the previous century, the area has seen little breakthroughs. In other words, we are still searching for ways to exploit our knowledge about the process to model into modern machine learning techniques. This is relevant for scientific fields that use mechanistic models, e.g. environmental sciences, cognitive science, embodied AI, modern robotics, emulation or surrogate modeling, etc. In this talk we will briefly overview the problem of regularizing and/or constraining regression with mechanistic knowledge.We will summarize some current results and highlight some remaining challenges.The presentation will be illustrated with examples and applications. BIO: Juan Pablo Carbajal was born in Salta, Argentina. He completed his studies in physics at the nuclear facilities in Instituto Balseiro, in Bariloche, Argentina and oriented his carrier towards applied physics and complex systems, completing his Master degree in this topic. After two years of applied research in Industry (Centre for Industrial Research, Tenaris, Argentina) he moved to Switzerland for his PhD at the Artificial Intelligence Laboratory of the University of Zürich. Soon after his graduation he spent two years as a postdoc at the Machine learning laboratory of the university of Ghent, Belgium working on physical analog computing using memristors. Since 2016 he has been working as a postdoc researcher at Eawag, applying machine learning techniques together with mathematical modeling methods to accelerate hydrological simulators. In 2017 he is part of a collaboration with the Institute for Energy Technology at HSR with the aim of accelerating CFD models used for wind turbine optimization. Besides his academic activities Juan Pablo leads the non-profit Dwengo Helvetica promoting robotics and science among teenagers.
