Geoffrey Hinton du Canadian Institute for Advanced Research et de l’Université de Toronto donnera une conférence le 23 octobre de 15hs à 17hs dans la salle DS-1950 (Métro Berri-UQAM, pavillon De Sève)
Abstract : Systems that adapt need to explore the space of possible adaptations. I will describe several biological examples where a system that adapts slowly avoids a lot of slow exploration by using a system that adapts much more rapidly to create a search space that is much easier for the slow system to explore. For example, learning changes the search space for evolution, thus allowing learning organisms to adapt their DNA much faster than organisms that do not learn. Similarly, reactions that happen at a sub-millesecond timescale in mass-spring systems change the search space for neural motor control. I will then argue that biological adaptation is entering a new phase in which computers will greatly accelerate the speed at which we learn by using rapid optimization to change the search space for learning. A designer, for example, will be able to explore the space of objective functions for evaluating designs rather than exploring the space of designs. Computer optimization is the next term in the series : Evolution, development, learning ...
Geoffrey E. Hinton’s Biographical Sketch