Uber: Tooling is a critical part of AI development and deployment

Uber employs thousands of machine learning models to inform all aspects of its business, according to chief scientist Zoubin Ghahramani. He revealed this tidbit during a session at VentureBeat’s Transform 2020 summit, during which he spoke about Uber’s use of AI and internet of things (IoT) technologies at the edge and in datacenters around the world.

Contrary to popular belief, autonomous vehicles aren’t the top driver of AI and machine learning at Uber, according to Ghahramani. (Uber’s Advanced Technologies Group has been developing and testing self-driving cars for passenger pickup since 2015.) Rather, the bulk of the company’s algorithms are designed to handle natural language interactions across Uber’s mobile apps and to detect fraud and other issues. In May, for example, Uber rolled out an AI system to verify drivers are wearing masks in accordance with the company’s pandemic health and safety policies.

Some algorithms are better suited to on-device edge processing than processing in the cloud, Ghahramani says. In some parts of the world, internet-based solutions are far less reliable — if they can be deployed at all. For systems like the kind responsible for identifying glare, blur, and truncation from photos of driver documents and identification, Uber uses “very small” mobile-optimized models that work in real time.

These and other models — both online and offline — are served by Michelangelo, Uber’s internal platform that enables teams to build, deploy, and monitor AI at scale. Michelangelo helps track model performance over time, providing transparency to engineers and executives, Ghahramani says. And it affords visibility into Uber’s data pipeline, allowing data scientists to spend time tracking and ensuring data quality.

Operationalizing AI

When asked whether Uber’s initial public offering in May 2019 changed its approach to AI, Ghahramani said the company shifted its focus from longer-term research to nimbler approaches that can respond to shocks like the pandemic. In April, the company said ride-hailing requests had dropped 80% globally. That same quarter, revenue from restaurant food deliveries rose by more than 50% year-over-year.

“We’re focused on showing return on investment. We try to ruthlessly prioritize the value of what we create,” Ghahramani said. “AI and machine learning is not magic — it’s as good as the data that you have, the tools that you use to extract value from that data, and the people that are operating those tools.”

One of these tools is Ludwig, a library built atop Google’s TensorFlow that’s used internally at Uber to train models without code. Others include Plato, a conversational AI development suite; Piranha, a tool that automatically deletes stale code; Manifold, a visual tool for debugging AI; and Neuropod, an abstraction layer intended to unify disparate frameworks like TensorFlow and Facebook’s PyTorch. All are available in open source.

“You have to invest in open source — just embrace it,” Ghahramani said. “It’s just the way people do things.”