Machine learning is all about inference! Given some data we would like to extract useful information, e.g., to predict that the given animal on an image is a dog or cat. Traditional machine learning techniques have been shown to work effectively in practice, but they also have their pitfalls: it's hard to directly incorporate domain knowledge, and they are often too confident about their results. Probabilistic programming languages (PPLs) augment traditional programming languages like Python, with first-class probabilistic constructs like stochastic variables, distributions, sampling and conditioning. Domain knowledge can be incorporated using priors, and predictions are specified as probability distributions over possible outcomes, thus quantifying uncertainty. The power of PPLs lies in the automation of Bayesian inference, making it now available to a wider range of programmers than before.
My talk will focus on providing an introduction to PPLs using the Pyro framework for Python, developed by Uber and the Linux Foundation. I will discuss how to use Pyro using example models like Bayesian logistic regression and Gaussian mixture models. I will summarize the available techniques for inference, highlighting their advantages and pitfalls. Finally, I will discuss practical applications of PPLs, both in the context of my work at Skanned.com, but also in the context of research and industry in general.
MetadataTo be recorded?: Yes
URLs for Why the fork should I care about probabilistic programming?
No URLs found.
No instances scheduled yet