Machine learning (ML) and deep neural networks are the systems that make it possible to perform voice searches on phones, translate webpages or street signs on the fly, beat human Go masters, or search Google Photos for pictures of cats and actually get useful results. They’re also the key drivers behind artificial intelligence (AI), which is why Google has announced a new, dedicated research group focused on ML.
Called Google Research, Europe, the new research group isn’t housed in the company’s biggest office in Mountain View, California, as many might expect it to be, but in the Swiss capital, Zurich, instead.
“Zurich is already the home of Google’s largest engineering office outside the US, and is responsible for developing the engine that powers Knowledge Graph, as well as the conversation engine that powers the Google Assistant in Allo,” says Emmanuel Mogenet, the Google engineer who’s heading the new research group. Allo is Google’s new messaging app that’s able to provide useful, contextual information based on the contents of a conversation — like suggesting a restaurant in a particular area when you ask a friend if they’d like to meet for dinner.
The new research group has three primary research focus areas: machine learning (and deep learning), natural language processing and understanding, and machine perception.
Google also hopes its new ML-focused research group will serve as a magnet for the many talented individuals Europe’s top-tier engineering faculties produce. “Europe is home to some of the world’s strongest technical universities, many of which focus on similar things,” Mogenet says. “This makes Europe the perfect place to build a top-notch research team.”
Mogenet says it’s the developments Google’s made in ML in recent years that have made services like Smart Replies in Inbox — which offers users suggested responses to emails — and Google Photos image search possible. It’s also changed the way search works.
“When you did a Google search five years ago we used to query specific terms and use keyword matching,” Mogenet explains. “Now we truly understand the question.”
It’s why you can ask Google voice search about a specific city, then ask what the tallest building is in that city, and then ask something like “How tall is it?”, and receive the right answers all the way down the chain. ML has enabled a degree of contextual awareness that was previously impossible.
“Historically, computers have been very good at number crunching and storing large amounts of data,” Mogenet says. “But they are really bad at doing things a small child could — like natural language, walking or identifying objects in images.
“The reason humans understand natural language is they understand the world around them because they’ve experienced it. Their senses have imparted certain information, from which they’ve learnt obvious things about the world. We’re teaching computers the same way and creating a ‘common sense knowledge database’ using images, video and other content.”
Teaching computers to understand the world is a slow process, though, and despite advances in computer vision of the sort that powers Google Photos, it’ll be some years yet before we see a computer than can understand that cows don’t fly for reasons other than because someone put that piece of information in a database. For the AI alarmists, this should come as some consolation. Computers may be getting smarter, and the ways we teach them may be changing, buy Skynet’s still a long way off.
Craig Wilson was at the Zurich launch event this week courtesy of Google.