The Google Brain project has been working on the tech giant's artificial-intelligence capabilities for years. But more recently, the company has become much more vocal about it. That's partly because machine learning is much closer to being one of the products that Google sells – and a feature of the products that already make up its revenue. In roughly the past two years, the progress in advancements of its machine-learning products "is more than we've seen in 10 years," Philipp Schindler, Google's senior vice-president and chief business officer said. " … It's just going vroom. It's mind-blowing."
Mr. Schindler sat down with The Globe and Mail during a recent visit to Toronto.
Explain to me how machine learning is affecting your business in ways it didn't before.
On search, we use RankBrain to improve our search quality. Categorizing YouTube videos. Google Photos – two years ago, the computer could not recognize a cat from a horse. Now, you go into the search box, and it shows me all the people I take pictures of most frequently. I've never tagged those people. It's my kids, my wife, my parents. I can sort for my three-year-old; I've never tagged my three-year-old in my pictures. This is the machine learning that does it. I can type in anything – sunset, hugs, horses – and it would show me the pictures based on visual identification.
A lot of people know machine learning as a buzzword, but don't really understand it. What are its concrete applications?
Machine learning is not a crazy thing that humans could not do before. Were you able to recognize your kids before? Yes. But were you able to instantly pull out, out of thousands of photos, every photo with your kid in it? No. Was a doctor in the past able to identify skin cancer? Of course. But you can [train] a machine-learning algorithm on millions of [photos of skin cancer]. No doctor in the future will dare to run a skin-cancer analysis without the help of a machine-learning algorithm.
As a fail-safe.
Exactly. Machine learning is not this crazy thing that will put people out of work. It will do things at a completely different scale. Machine learning is pattern recognition at an incredible scale and quality.
How does this impact advertising, which is the bulk of Google's business? Are there examples of uses of machine learning in your ad business, where you weren't using it a couple of years ago?
YouTube is a good one. We have 400 hours of video being uploaded every second. How do you categorize them? It's humanly not possible any more. You're aware of the issues we had around brand safety. How do we get a better understanding of the type of videos we don't want to make available for [advertising next to them] – or even not have at all on the platform? Machine-learning systems are going through this.
How much work still needs to be done to guarantee brand safety on YouTube?
We have taken this super seriously. As seriously as any issue we've ever faced. We've made massive progress. We've dramatically increased the investment on engineering resources, on brand-safety machine learning. And we have also hired a significant number of people: human raters, to create the data signals that the machine can learn on. If something still goes wrong, we've changed the escalation process. The safety aspect, I think we have that under control. It will never be 100-per-cent perfect. You can always have a small mistake popping up somewhere. But the brand-suitability piece, that's the biggest discussion we have at the moment with most advertisers. Different brands have very different perceptions – everybody agrees that we don't want terrorist content, that's obvious – but what is acceptable for one brand is completely different than another brand. So, what type of filters do you have? How do you make it easy? Where do you draw the line?
Are there other examples of how machine learning affects advertising?
On Android [mobile devices], roughly 20 per cent of searches are voice searches. If you think about how you interact with digital devices today, it's probably still a bit clunky that you still have to type on a keyboard into a search box. The next evolution you will see is toward [voice-activated services]. You might have one on your phone, in your car, in your television set, in a wearable, in your Google Home device. They will all be connected, it will be the same assistant. The more the assistant understands you and your needs, the better the answer or the suggestion quality. That has massive implications on the advertising side. Forced, non-targeted, bad-quality advertising frankly doesn't have a future on any of those assistive services. The good news is that the number of interactions of consumers with digital devices will significantly go up, which is a great opportunity for business.
What about beyond advertising?
The next thing we're going to do is take all the machine-learning intelligence we've developed it and offer it as a service to every other industry out there. It's very similar to the cloud business – we and other players had to develop very complex infrastructure and then we offered it as a product. We're going to do the same with machine learning. Every insurance company, every bank, every retailer, whoever wants to use some of this magic will be able to access our APIs [application program interfaces] to use it. It will not be the normal adoption curve – where somebody is like, "I have to hire the people, create the infrastructure, train them." No, no, no.
This raises privacy issues. People might be unsure whether they want Google to recognize their child's face in thousands of photos, for example. What are you doing to address that?
We cannot promise the user, "I don't need data to give you a really good answer with an assistant." It doesn't work. I cannot say to the user, 'I don't need to understand your voice model better' [to make voice recognition work.] … We have to fully recognize those fears, of course. The "my account" interface gives you full control over your information at Google. It's actively being used, so it's not some hidden feature. Remember, in all those applications when we talk about machine learning, we talk about anonymized data. The personal identity of the person is irrelevant for us. But we give the user full control. It's all about transparency and control. That's the contract we have to have with the user in the future.
This interview has been edited and condensed