2016: the year AI came of age A nice article on AI in The Guardian

Looks like 2017 will be the turning point in AI commercialization and this article sums up the current situation quite well.

However the really important bit to grasp is where it mentions "moving from explicitly taught systems to implicitly learning ones". This will bring the real acceleration by using 'big data' that it collects in the process. It is not only scary in terms of impact on jobs in the service sector but also in terms of what it can lead to in the 'very near' future.

Google and Amazon brought AI into the home and DeepMind built a computer that could outsmart humans at Go. Will 2017 hold similar advancements?

We'll see...

Alex Hern

@alexhern

Wed 28 Dec 2016 05.00 EST Last modified on Sun 16 Jul 2017 16.45 EDT

Over the course of 2016, artificial intelligence made the leap from “science fiction concept” to “almost meaningless buzzword” with alarmingspeed.

Everything has AI now. Period-tracking app Flo “uses a neural network approach” to deliver “high period forecast accuracy”; food delivery app Just Eat launched a chatbot that “sees AI integrated into the ordering experience to ensure that customers receive the best, round the clock support and service”; restaurant guide Borsch “uses artificial intelligence to help people discover the yummiest dishes around”.

But unlike many buzzwords before it, from “big data” to “blockchain”, artificial intelligence’s transformation into venture capitalist-catnip doesn’t signify the end of anyone serious using the term themselves. In fact, 2017 looks like it could be the most important year yet for the technology: AI will butt up against not only what is possible, but also what is desirable for the first time.

Like many futures, the AI revolution feels interminably slow to live through, and will feel like it happened in an instant in hindsight. The first pivotal year was 2011. That was when Apple’s Siri hit iPhones, introducing the world to the first major “virtual assistant”. It was also the year the Google Brain project was instituted: the search engine’s blue-sky research team aimed to address as many tasks as possible through neural network-based learning, the computational technique that has come to define what we mean by artificial intelligence.

Five years on, and neural networks have already begun to enable tech which seemed impossible back then. Google and Apple have applied them to their photo apps to let users search through their pictures for images of “dogs”, “cars” or, in Google’s case, “Christmas”, based on what the algorithms see in the images. That machine vision technology is also the basis of the self-driving car efforts from Google’s sister firm Waymo. Oh, and an entirely different neural network is probably the world’s best player at the ancient boardgame Go.

That victory, from Google subsidiary DeepMind, was one of the last remaining milestones for a machine to reach. Go is so complex that, as recently as 2014, many thought it would be another decade until an AI could approach the skill of a human player. That was what made it so appealing for DeepMind to tackle.

There’s one remaining milestone that the London-based research lab is interested in chasing, according to co-founder Mustafa Suleyman, and it’s a big one: instant voice-to-voice translation. The company has slowly been assembling the pieces for a while, with Google already rebuilding its translation service around a neural network-based approach, and DeepMind creating a whole new way of synthesising speech it calls WaveNet, but there are still a host of other problems to be overcome before the babel fish becomes a reality.

Which is not to say that 2017 won’t be a groundbreaking year for AI. The biggest effect will be the step change in the amount of data which companies such as Google and Amazon have access to. When Google released its voice-controlled, AI-powered smart home device, Google Home, in 2016, it already impressed some with its abilities. But, says Fernando Pereira, who leads Google’s natural language understanding projects, that’s only the start.

Now that millions of people have Google Home in their living room, the company can analyse every natural language query it starts getting from all of them, giving it far more data to crunch than it could ever get from its testers. “You can start doing machine learning on that,” Pereira told tech site Backchannel. “You can move much faster; you can accelerate the process of getting deeper and broader in understanding. This 2016-to-2017 transition is going to move us from systems that are explicitly taught to ones that implicitly learn.”

This is the story Google wants to tell of machine learning: an acceleration, turning the coming year into an inflection point, the instant that machine learning became good enough to start trusting.

It’s certainly one possible outcome of the next year, although it’s not yet clear whether Google will be the one to deliver on it; Amazon has been keeping pace with its own Alexa assistant, for instance, while others including Facebook, Microsoft, IBM and Baidu have been trumpeting their own machine-learning successes.

But the other possibility is that, as machine learning steps out of the shadows and companies ask for ever more data to train their algorithms, the backlash begins. Already, Google faces competition from other companies over how much of your life it wants to manage.

That happens implicitly, in the difference between Google Home and Amazon’s Echo: the former integrates tightly with your Google account, reading emails, notes and calendar events to keep up to date with your life, while the latter takes a more hands-off approach, only linking with what it’s told and generally attempting to be responsive, rather than proactive.

It also happens more explicitly in the way Apple has decided to weigh in against its rival. The company, freed from the need to data mine everything by its old-fashioned “sell things for money” business model, has been proudly demonstrating approaches to AI which don’t need a central repository of harvested data to learn or work. That includes its machine vision approach, which scans users’ photo libraries on device, rather than on the cloud, and its research into “differential privacy”, a technological approach to machine learning which allows the company to learn from data in aggregate while never having access to the information of specific users.

Of course, there is a third option: that neural network-based machine learning will instead prove to be a technology like any other, useful in some areas, useless in others, and eventually doomed to be rendered obsolete in turn by a future innovation. We’re already seeing some of the downsides, in the eternal craving for more data, in the processing power required to actually learn, and in the opacity of the models that result. One day, those downsides will outweigh the up, and the world will move on. But for now, there’s still a world of possibility.