Introduction to DeepHyper

Abstract:  Machine Learning algorithms are very powerful, but their performance strongly depends on choosing the suitable hyperparameters.  This choice is not easily made, even with expert knowledge.  The hyperparameter space is usually very complex, mixing continuous, integer and categorical variables.  One needs a systematic way of optimizing algorithm performance over this space, which is the goal of the DeepHyper framework that I will introduce in this talk.  We will then see that our optimization has a very nice byproduct:  a trained ensemble of learners.

Speaker’s Bio: Caio Alves has a Ph.D. in mathematics from the Federal University of Minas Gerais, in Brazil. He worked on discrete probability theory, focusing on stochastic processes on graphs.  Since joining the Oak Ridge National Laboratory in November 2022, he has worked on probabilistic modelling, tensor factorization on semiring spaces, and ensemble learning.