Automatically Constructing Models And Automatically Explaining Them Too
How could an artificial intelligence do statistics? It would need an open-ended language of models, and a way to search through and compare those models. Even better would be a system that could explain the different types of structure found, even if that type of structure had never been seen before. This talk presents a prototype of such a system, which builds structured Gaussian processes regression models by combining covariance kernels to build a custom model for each dataset. The resulting models can be broken down into relatively simple components, and surprisingly, it's not hard to write code that automatically describes each component, even for novel combinations of kernels. The result is a procedure that takes in a dataset, and outputs a report with plots and English descriptions of the different types of structure found in that dataset. I'll also talk about advances in black-box stochastic variational inference methods, which have the potential to open the door to even broader model classes.
David Duvenaud is a postdoctoral researcher in the Harvard School of Applied Sciences and Engineering. He obtained his doctorate in machine learning at the University of Cambridge. His research has focused on probabilistic models of functions, with applications to forecasting, numeric computations, and deep learning. David previously worked on machine vision at Google research, and co-founded Invenia, an energy forecasting and trading firm.