Case Study: Allen Institute for Artificial Intelligence improves reproducible research with Google Cloud and Google Kubernetes Engine

AI institute uses Google Cloud to solve everyday challenges

The Allen Institute for Artificial Intelligence (AI2) uses Google for Education’s Google Cloud platform to support researchers who need to run many experiments, track results, and improve reproducibility. AI2’s challenge was reducing the infrastructure and collaboration overhead that slowed experimentation, especially for complex deep learning workflows with many parameters and dependencies.

Google for Education helped AI2 build Beaker, a Kubernetes-based platform running on Google Kubernetes Engine and Google Cloud to automate experiment tracking, sharing, and infrastructure management. The move improved stability and transparency, made it easier to scale and customize workloads, and let researchers focus on science instead of operations. AI2 reported more than 60,000 experiments and nearly 190,000 GPU compute hours in 2018, plus 33,000 experiments and 60,000 compute hours in the first two months of 2019.

Allen Institute for Artificial Intelligence

Marc Millstone

Engineering Lead