UCPH Statistics Seminar: Melih Kandemir
Speaker: Melih Kandemir from the University of Southern Denmark
Title: Adaptive Intelligence with Neural Stochastic Processes
Abstract: Most artificial intelligence applications employ a training period that ends prior to the time the learned model starts to be used in the real world. This common practice builds on the assumption that the observations available to the model during training time are accurate representatives of all possible scenarios it can encounter during its whole life time. This assumption is in sharp contrast with the perpetually changing environments of interactive agents, such as robots. A model can both improve its predictions on its assumed environment and adapt to environmental changes if it is allowed to update itself while it is in use. My newly established research lab works on closed-loop training of interactive agents, which we call as the “adaptive intelligence” problem. In this talk, I will introduce life-long inference of neural stochastic processes via bandit algorithms as a methodological framework well-suited for the adaptive intelligence problem. I will also summarize the outcomes of our recent efforts to solve it.