You are here

It's time to talk about irregularly-sampled time series

David Duvenaud
University of Toronto

MIA Special Seminar: It's time to talk about irregularly-sampled time series
Most biological data is sampled at irregular intervals, but most machine learning time-series models require regularly-spaced observation times. I'll introduce the pros and cons of different ways to deal with this mismatch. Then I'll describe the options for building continuous-time models: autoregressive vs. latent variable, and deterministic versus stochastic. Finally, I'll show new time series models currently being developed, based on latent variables described by ordinary or stochastic differential equations, and their application to an intensive care unit dataset.