Skip to main content
Uber logo

Start ordering with Uber Eats

Order nowOrder now

Start ordering with Uber Eats

Install the appInstall the app
AI, Engineering

Announcing a New Framework for Designing Optimal Experiments with Pyro

May 12, 2020 / Global
Featured image for Announcing a New Framework for Designing Optimal Experiments with Pyro
Figure 1. The Bayesian logistic regression model for the working memory experiment encodes how the probability of a correct response varies with the length of a list of digits.
Figure 2. This animated graphic illustrates the sequence of steps in an adaptive experiment designed to efficiently probe human working memory capacity.
Figure 3. The posterior for an uninformative experiment is barely different from the prior.
Figure 4. The posterior for an informative experiment is substantially narrower than the prior.
Figure 5. The computed EIG varies considerably as a function of the design; at this stage of the experiment, we expect sequences of digits of length 7 to be most informative.
Figure 6. As more data is collected, the posterior over working memory capacity sharpens around the value of 5.5.
Figure 7. The uncertainty in the posterior after a series of experiments guided by OED is less than that for heuristically chosen designs.
Adam Foster

Adam Foster

Adam Foster was a research intern at Uber and is currently pursuing a Ph.D. at the University of Oxford.

Martin Jankowiak

Martin Jankowiak

Martin Jankowiak is a senior research scientist at Uber whose research focuses on probabilistic machine learning. He is a co-creator of the Pyro probabilistic programming language.

Posted by Adam Foster, Martin Jankowiak