Poster
BatchBALD: Efficient and Diverse Batch Acquisition for Deep Bayesian Active Learning
Andreas Kirsch · Joost van Amersfoort · Yarin Gal

Thu Dec 12th 10:45 AM -- 12:45 PM @ East Exhibition Hall B + C #1
We develop BatchBALD, a tractable approximation to the mutual information between a batch of points and model parameters, which we use as an acquisition function to select multiple informative points jointly for the task of deep Bayesian active learning. BatchBALD is a greedy linear-time $1 - \frac{1}{e}$-approximate algorithm amenable to dynamic programming and efficient caching. We compare BatchBALD to the commonly used approach for batch data acquisition and find that the current approach acquires similar and redundant points, sometimes performing worse than randomly acquiring data. We finish by showing that, using BatchBALD to consider dependencies within an acquisition batch, we achieve new state of the art performance on standard benchmarks, providing substantial data efficiency improvements in batch acquisition.

Author Information

Andreas Kirsch (University of Oxford)

AIMS DPhil at University of Oxford, currently in 2nd year/4 (first paper: http://batchbald.ml) DeepMind: performance research engineer for 1 year Google: software engineer for 3 years MSc CompSci, BSc CompSci, BSc Maths at TU Munich

Joost van Amersfoort (University of Oxford)

I am currently pursuing a PhD at the Unversity of Oxford under supervision of Professor Yarin Gal (in OATML) and Professor Yee Whye Teh (in OxCSML). I am interested in variational inference (for example done through the reparametrisation trick, dropout and reversible models) and its applications such as uncertainty estimation, active learning and generative modelling. Previously, I was a research engineer at Twitter Cortex, where I worked on large scale video models and recommendation systems. I did my masters at the University of Amsterdam, where I worked with Dr. Durk Kingma and Professor Max Welling. I've also done an internship at Facebook AI Research (FAIR).

Yarin Gal (University of Oxford)

More from the Same Authors