Now the taxi company Uber is getting into the AI business.
Yesterday it announced it was buying an AI startup called Geometric Intelligence for an undisclosed sum.
Why would a taxi company need an AI division? Two main reasons.
The first is that Uber is now a rather big company. It’s worth $70bn but it’s still reliant on other companies’ technologies, such as Google Maps.
When a business is worth nearly $70bn, and is increasingly in competition with Google, it makes sense to be self-reliant. Uber doesn’t want to be in a position where Google can shut down its business at the flick of a switch by, for example, taking away Google Maps. That’s why Uber has been building out its own technology “stack” – it’s investing $500m in building its own maps, for example. The AI division is part of that overall strategy.
That’s the general reason why Uber is doing AI in-house. But, specifically, why does a taxi company need an AI division?
This next bit is more speculative. We know Uber is working on some knotty engineering problems, such as self-driving taxis and efficient carpooling. State of the art machine learning will help with those.
But you have to remember, AI can be used to solve many different types of problem. It’s what they call a general purpose technology. And it’s getting better all the time.Perhaps it’ll be used to help come up with more efficient carpooling routes on day 1, but in five years who knows what AI will be capable of. And in five years’ time Uber doesn’t want to have to go cap-in-hand to Google, looking to licence its AI technology.
Slow on the uptake
Geometric Intelligence, Uber’s new acquisition, is an interesting company.
It takes a different approach than most of its competitors. The competitors train computers using a technique called “deep neural networks”. For example: they train a computer to recognise a cat by showing it half a million pictures of a cat; then ask it to pick out cat pictures itself; then correct it. In this iterative way the computer figures out what a cat looks like.
This approach works for lots of different types of problems. It’s responsible for most of the AI breakthroughs in recent years. But its drawback is that you need loads of a data – half a million labelled pictures of a cat – in order to train it. Only a few companies in the world have access to that amount of data.
In other words, computers are slow on the uptake. The next step is to come up with smarter learning algorithms that don’t rely on brute force and half a million examples to solve problems. A child, for example, can learn what a cat looks like from 10 examples. The holy grail of AI is to come up with a learning algorithm that efficient.
That’s the problem Geometric Intelligence has been working on – teaching computers to learn more quickly and efficiently. It was founded by Gary Marcus, a famous linguist, and Zoubin Ghahramani, a University of Cambridge mathematician. Speaking to Wired, Ghahramani said the technology is a hybrid of two different AI approaches, “deep learning”, and “rules-based”.
“If you combine some of the ideas in ruled-based learning with ideas in statistical learning and deep learning, then you can get the best of both worlds.
“If there is an obvious rule—or even if it’s not so obvious—they will eventually catch on to that, and they’ll generalize to new situations. But they can pick up statistical patterns from lots and lots of data as well.”
In other words: programming computers to learn things more like people do.