SMBp Group Meeting: Charles Margossian (CCM)

America/New_York
3rd Floor Classroom/3-Flatiron Institute (162 5th Avenue)

3rd Floor Classroom/3-Flatiron Institute

162 5th Avenue

40
Description

Speaker: Charles Margossian  (CCM)

 

Topic: From high-performance algorithms to high-performance modeling


Abstract: Probabilistic programming languages such as Stan allow practitioners to deploy high-performance algorithms and apply them to many modeling problems. For example, the broad use of Hamiltonian Monte Carlo (HMC) can be attributed to its implementation in Stan and other statistical software. But from HMC's inception in 1987 to Stan's first release in 2012, much technological innovation had to happen. I will focus on three such innovations: (i) self-tuning algorithms, (ii) diagnostics, and (iii) automatic differentiation. Arguably, the story of HMC provides a template for making other promising algorithms wildly usable---including other types of MCMC and variational inference algorithms---, although we should also acknowledge persistent gaps in HMC's current workflow. Applications in pharmacometrics and epidemiology will be used as running examples.

The agenda of this meeting is empty