In 2008, Antoine Blondeau co-founded artificial intelligence (AI) startup Sentient Technologies, with a team that had worked on laying the foundations of the technology that would become Apple's Siri. The 60-employee company has raised $143 million, making it the world's most funded AI company. Its goal? Using AI to solve any problem - from financial trading to deadly diseases.
Sentient's core technology is a deep-learning algorithm running on a network of more than two million computers worldwide. On such a wide computational platform, multiple AI systems constantly vie to pick the best choice, evolving over time as the worst performers are culled out. According to Blondeau, scalability and evolution are key to AI's future. For the time being, Sentient's applications are mainly in the financial field. The company already provides automated financial services that become gradually more effective at placing investments. More recently, the company partnered with e-commerce website Shoes.com, reinventing its system to advise shoppers by learning their fashion tastes.
Sentient is also working on an AI nurse, a system that would monitor a patient's vital signs to predict when they might get a certain condition or disease. The idea has already been tested in a joint trial that Sentient carried out with the Massachusetts Institute of Technology: the AI nurse managed to predict within 30 minutes if a patient would get sepsis by looking only at their blood pressure. WIRED caught up with Blondeau to discover what AI has in store for us.
WIRED: You say one of Sentient's advantages is the ability to scale. What does that mean in an AI context?
Antoine Blondeau: For us, scalability means the ability to take our AI construct and extend it within many computers - two million computers, in fact1. If you are tackling a large problem, you need more computational power, that's the basic assumption. That's why you want to be sure that the AI layer at the top of the computers can scale. And a lot of our talent resides in the ability to architect the AI system and deconstruct any problem in a way that allows you to send jobs everywhere. Another critical skill is aggregating the plural outcomes into one, meaningful, whole solution.
Sentient relies on this firepower to make AI systems evolve. How does the evolution process play out, concretely?
OK, let's say you want to create AI financial traders. So in the very first phase of evolution that you create, you generate an enormous amount of random traders. They'll look at all the data points in the universe and they'll start to construct rules, very simple rules in the beginning. Something like: "If I see this and that, and if something else happens, then I'll do this."
You can use scale, your enormous computer network, in order to create trillions and trillions of such traders. And, as those traders are created totally randomly, and the rules are created randomly too, at first they are all going to be rather crappy. But some of them will be better than others. So you take the ones that achieve the best results, which make the best predictions, and you use their "DNA", their rules, in order to create a second generation of traders. Even then, though, you inject some randomness… and so forth. This process goes on all the time. And it is very robust because you create species that are inherently very apt, as they have seen an enormous amount of things and their "DNA" has been tested. But you also get diversity, as there are many possible solutions.
What is the critical technology that allows you to make this?
There are at least three. First, we've developed an architecture that enables us to identify online underutilised or unused computers: this enables us to talk to game centres, universities, call centres and enterprises, so they give us their spare computers. On the evolution side we have relied on evolutionary algorithms. Third, we created an online learning platform to teach AI systems to recursively reiterate processes and learn to recognise images, pictures, stock-market graphs and so on.
You say that over time AI systems would learn what the best rules to deal with a problem are. Can these systems, when they are evolved, teach us how to solve problems?
Well, one of the good things about evolutionary AI is that - if you know how to read it - you can actually see the rule sets2. In the case of traders, or of AI nurses (on which we are working, too), they are fairly complex beings: a trader may have up to 128 rules, each with up to 64 conditions. Same thing for an AI nurse. So, they are pretty complex systems and the interplay among these rules is not always linear. But if you spend some time on it, you can still understand what this thing is doing, because it's declaratory - it says what it is doing, in other words. So we can certainly take this and learn from this what works and what doesn't work when it comes to solving a certain problem. AI can teach people to make better decisions.
1. "It takes a lot of horsepower to do what we do," says Blondeau. Sentient's network uses 5,000 GPU cards in 4,000 sites.
2. Sentient is working with the Oxford Genomics Centre to understand why some people develop genetic diseases. "We're using simulated evolution, which takes days, to understand our own biological evolution, which took billions of years."
http://www.wired.co.uk/…/sentient-technologies-ai-nurses-ba…
Cecile G. Tamura
No comments:
Post a Comment