A company called Bonsai joins a movement to democratize machine learning. Get ready to build your own neural net
If you are a strong-armed NFL quarterback who reads defenses like genre fiction, a movie star whose name alone can open a film in China, or a stock picker who beats Buffet every time, congratulations: you are almost as valuable as a data scientist or machine learning engineer with a PhD from Stanford, MIT, or Carnegie Mellon. At least it seems that way. Every company in Silicon Valley — increasingly, every company everywhere — is frantically competing for those human prizes, in a human resources version of a truffle hunt. As businesses now realize that their competitiveness relies on machine learning and artificial intelligence in general, job openings for those trained in the field well exceed all the people in the world who aren’t locked up by Facebook, Google, and other superpowers.
But what if you could get the benefits of AI without having to hire those hard-to-find and expensive-to-woo talents? What if smart software could lower the bar? Could you get deep learning with a shallower talent pool?
A startup called Bonsai and an emerging class of companies with the same idea say yes. Brace yourself for the democratization of AI. It’s a movement that might eventually include millions of people — and, some say, billions.
Today, at the O’Reilly Artificial Intelligence Conference in New York City, Bonsai’s CEO Mark Hammond will give his company’s signature demo. (He will also announce a $6 million investment round, which is not so amazing considering venture capitalists have invested over $1.5 billion in AI startups already this year.) The demo involves duplicating one of the iconic achievements of an elite deep learning operation: DeepMind’s spontaneous mastery of the game portfolio of old Atari computers. Specifically, the familiar contest called Breakout, in which a paddle bounces a square-ish “ball” to erode a wall of glowing “bricks.” (The 1976 game was cutting edge in its time — Steve Jobs worked on it!)
DeepMind’s implementation sprung from its roster of world-class AI geniuses who trained a neural net to tackle a whole set of Atari games; the achievement was worthy of publication in a world-class journal. Bonsai’s version is the result of a shortcut. It starts with the company’s development system, sitting in the cloud. A single programmer, maybe one who never took an AI college course or even a MOOC in the subject, can outline a game, and the system will choose the appropriate learning algorithm to get the neural net underway. (The poor PhDs from DeepMind had to choose or write the algorithms themselves.) From there, the programmer lays out the concepts of the game in a few minutes—things like keep paddle on ball—and lets Bonsai do the neural networking, optimizing for high score. The resulting network will play a mean game of Breakout.
The Bonsai version of the game is only 37 lines of code. But that’s deceptive. When Hammond reveals what’s going on underneath the hood, he shows a graphic that illustrates how his system has built a complicated neural net worthy of one of Google’s machine learning ninjas. The programmer never had to deal with all that machine-learning stuff. Look Mom — no PhD!
It’s an eyebrow-raising stunt. “I’m not usually captured by demos,” says George Williams, a research scientist at NYU’s Courant Institute of Math. “But what Mark showed me was both plausible and just amazing. He’s captured where we are with machine intelligence and the tools we need to build the next generation of AI.”
Whether Bonsai itself will wind up leading this movement is not a sure thing. But Williams is right about the “where we are” part. The next step in the inexorable rise of smart computers is machine learning for (relative) dummies.
Bonsai began on a beach. Hammond, a former Microsoft engineer and developer evangelist, had been pondering artificial intelligence for a while. After leaving Microsoft in the early 2004, he worked at Yale in neuroscience; in 2010 he spent a brief period at Numenta, an AI startup led by Jeff Hawkins (a co-founder of the Palm handheld assistant), but left to start an unrelated company that he eventually sold off.
Then in 2012, Hammond was in Southern California visiting friends. His toddler son was tired, and the group trekked back to the car. While Hammond’s wife chatted with their friends, and his son was falling asleep in his arms, he ran a thought experiment. It began with a popular meme in the AI world: the concept of a “master algorithm.” As posed by University of Washington professor Pedro Domingos (in a great book by the same name), this not-yet-discovered machine learning technique would be a one-stop solution to all sorts of problems. Once scientists devise this algorithm, the thinking goes, we will be able to methodically artificial-intelligencize everything.
But Hammond saw a flaw in this thinking. Let’s say we found that master algorithm, he thought to himself as his 18-month-old son dozed in his arms. Who would implement it in the countless use cases that would arise? Currently, only adepts in machine learning are capable of wielding such tools. And there are far too many uses for the meager number of those people to address. We need a system, he concluded, that would lower the bar so that your garden-variety software developer could use those tools. This system wouldn’t require highly specialized computer scientists to train neural nets, but would allow programmers to teach systems how to produce the desired effect.
As Hammond refined his concept, he developed an analogy to the history of computer programming. Originally, computer jockeys had to painstakingly write code that directly addressed the raw hardware. Then coders adopted standard instruction sets, called assembly language, that sped the process up — but you still had to be a pretty hardcore programmer to master assembly language. The breakthrough came when engineers created the compiler—a translator that converted what were called easier-to-use “higher-level” languages (ranging from BASIC to LISP to current ones like Python and C) into assembly language. Only then could programming be broadened to allow relative novices to create powerful applications. Hammond argues that with tools like Google’s TensorFlow, AI is now in the assembly language era, which makes it easier for the scientists building neural nets, but still limits the field to those who really understand how those nets work. His idea was to provide the equivalent of a compiler, to really open things up.
He shared the idea with Keen Browne, a former Microsoft colleague who had recently sold his gaming startup to a Chinese internet company. The concept resonated with Browne, who had been frustrated trying to do deep learning using the popular tools available. “I’m a pretty smart guy,” he says. “I went to China and learned to speak the language. I programmed at Microsoft. But doing this was ridiculous.” He signed on to co-found Bonsai. (The name was chosen because those artfully stunted Japanese trees balance both the natural and the artificial. A bonus came when new internet domains let the fledgling company register the address bons.ai.)
Bonsai isn’t alone in addressing the scarcity of skilled AI scientists. Some of the bigger companies have figured that they will do training in-house to boost their everyday coders into masters of neural nets: Google has developed a host of internal programs, and Apple looks for traits in programmers that indicate they can pick up those skills without much difficulty. As mentioned before, Google also has publicly released TensorFlow, the software it uses to help its own scientists build neural networks. Other AI toolkits are also available in open source, and more will undoubtedly follow, some of them requiring less expertise than others.
As Hammond describes it, building a neural net with Bonsai has some key differences from the way the pros do it. Currently, you have to choose which tools are appropriate for the problem, a decision that requires experience and knowledge. According to Hammond, Bonsai figures that out for you. All you have to do is lay out the concepts you want to teach to the system.
So while experienced AI scientists will “train” a net by comparing its output to the desired results (for example, showing it pictures of dogs and rewarding outputs that show canine characteristics), Bonsai allows you to “teach” a system by breaking down the process into concepts. For the dog example, you might specify things like four legs, snout, big sloppy tongue hanging over the jaw. You give things a push, and Bonsai’s cloud-based “intelligence engine,” including its “Brain,” sorts things out.
This comes with a positive side effect: after traditional neural nets are trained, scientists often have no idea how they perform their magic, because those nets largely configure themselves, organizing the various concepts in their own inscrutable manner. But with Bonsai, the concepts a user specifies provide a Baedeker to the neural net’s thinking. “The software should not be a black box,” says Hammond. For instance, he explains, if you are programming a self-driving car and the vehicle doesn’t hit the brakes at the right time, you should be able to drill down at that given moment and find out what the system was thinking — kind of like when Amazon explains why it recommended a specific book for you.
One big question arising from Bonsai’s approach is whether all this abstraction will slow down performance and effectiveness. That’s typically what happens with compilers: programs that use them don’t run as quickly or efficiently as those written in assembly language, direct to the hardware. Also, it’s a stretch to think that a system that chooses which tools to implement will match the acumen of those PhDs who are supposedly no longer needed to build those nets.
“I think you always have some tradeoffs,” says Lila Tretikov, an AI scientist who was formerly head of the Wikimedia Foundation and has advised Bonsai. “It’s not exactly the same as if I had a team of PhDs. But I don’t know if it’s as important as just being able to do it.” Adam Cheyer of Viv also surmises that Bonsai code may not run as efficiently as software optimized for a specific system. “But it’s pretty darn good code, and lets you stay at a higher level of abstraction,” he says. Cheyer says that his own company, loaded with those valuable AI scientists, will likely not use Bonsai, except possibly as a tool to prototype something before implementing it the good old-fashioned way.
For his part, Hammond claims there is very little falloff that comes from using Bonsai. “It gets better performance at the end of the day,” he says. “It’s one of these things where the proof is in the pudding.” Now that the pudding is being served, we’ll see.
Bonsai has big plans over the next few months. Another thing it is announcing this week is a deal with the chipmaker Nvidia, to make sure that Bonsai customers can efficiently run their neural nets on that specialized hardware. The company will also make public its arrangement with Siemens TTB (tech to business) Center, which has been testing Bonsai for the past few months in industrial automation and control.
Bonsai itself is trying to crack problems that even the powerhouses of AI haven’t resolved. “We are working on a lot of games,” says Hammond, explaining that games are proxies for a lot of the key problems Bonsai hopes to address. “There are certain classes of games that have not been solved, even by DeepMind. While they trained lots of other Atari games besides Breakout, they never got their system to successfully play Pac-Man.”
But the bigger story is how Bonsai fits into the movement to put AI into the hands of people who have not had specific training in it. We can expect higher-level tools to become increasingly powerful and, ultimately, ubiquitous. Will we really get to the point where every human in the world trains and uses AI? Put it this way: a lot of smart money is betting on it.
“We’ll have analysts in the cloud,” says Bottlenose CEO Spivack. He says we could call on these virtual consultants for “personal decision support,” to answer questions like, “Which college should I go to?” The cost will be nominal, if not free. “There will be no excuse to make poor decisions just because you can’t afford AI,” he says.
Maybe we’ll even get to the point where AI can conquer Pac-Man. Bonsai isn’t there yet. “We are working on it,” says Hammond. “Nothing to announce on that yet.”
Source @ BackChannel.
Visit Bosai home page @ https://bons.ai