Speaker: Arya Pourzanjani

 

Title: Some Things You Won't Learn in Your Machine Learning Class

 

Abstract: Over the last few years, the field of machine learning has made tremendous strides in using big data to fit models. At the same time, tantamount issues previously addressed in the “small data” regime, such as independence, sampling, and uncertainty have mostly been forgotten in the big data flurry. I will briefly talk about the cultural differences between machine learning and classical statistics, and illustrate, using examples from medicine, why we must consider important concepts previously addressed by the field of statistics, when working with big data. Specifically, I’ll explain how researchers and practitioners have been using “big data” to fit “small models”, and illustrate why it is important to use “big models”, especially with “big data”. I will then illustrate with examples, why in the big model regime, fitting models via optimization fails, and how these failures have manifested in current machine learning and specifically neural network research. I'll explain how and why sampling is a more robust alternative to optimization, and present the Metropolis-Hastings algorithm,and its newer, even more efficient cousin, the Hamiltonian Monte Carlo (HMC) algorithm. I will illustrate how HMC allows us to elegantly fit our models in a simpler and more automatic fashion than optimization.