by Cade Metz, Wired.com
Igor Mordatch is working to build machines that can carry on a conversation. That’s something so many people are working on. In Silicon Valley, chatbot is now a bona fide buzzword. But Mordatch is different. He’s not a linguist. He doesn’t deal in the AI techniques that typically reach for language. He’s a roboticist who began his career as an animator. He spent time at Pixar and worked on Toy Story 3, in between stints as an academic at places like Stanford and the University of Washington, where he taught robots to move like humans. “Creating movement from scratch is what I was always interested in,” he says. Now, all this expertise is coming together in an unexpected way.
Born in Ukraine and raised in Toronto, the 31-year-old is now a visiting researcher at OpenAI, the artificial intelligence lab started by Tesla founder Elon Musk and Y combinator president Sam Altman. There, Mordatch is exploring a new path to machines that can not only converse with humans, but with each other. He’s building virtual worlds where software bots learn to create their own language out of necessity.
As detailed in a research paper published by OpenAI this week, Mordatch and his collaborators created a world where bots are charged with completing certain tasks, like moving themselves to a particular landmark. The world is simple, just a big white square—all of two dimensions—and the bots are colored shapes: a green, red, or blue circle. But the point of this universe is more complex. The world allows the bots to create their own language as a way collaborating, helping each other complete those tasks.
All this happens through what’s called reinforcement learning, the same fundamental technique that underpinned AlphaGo, the machine from Google’s DeepMind AI lab that cracked the ancient game of Go. Basically, the bots navigate their world through extreme trial and error, carefully keeping track of what works and what doesn’t as they reach for a reward, like arriving at a landmark. If a particular action helps them achieve that reward, they know to keep doing it. In this same way, they learn to build their own language. Telling each other where to go helps them all get places more quickly.
As Mordatch says: “We can reduce the success of dialogue to: Did you end up getting to the green can or not?”
To build their language, the bots assign random abstract characters to simple concepts they learn as they navigate their virtual world. They assign characters to each other, to locations or objects in the virtual world, and to actions like “go to” or “look at.” Mordatch and his colleagues hope that as these bot languages become more complex, related techniques can then translate them into languages like English. That is a long way off—at least as a practical piece of software—but another OpenAI researcher is already working on this kind of “translator bot.”
Ultimately, Mordatch says, these methods can give machines a deeper grasp of language, actually show them why language exists—and that provides a springboard to real conversation, a computer interface that computer scientists have long dreamed of but never actually pulled off.
These methods are a significant departure from most of the latest AI research related to language. Today, top researchers typically exploring methods that seek to mimic human language, not create a new language. One example is work centered on deep neural networks. In recent years, deep neural nets—complex mathematical systems that can learn tasks by finding patterns in vast amounts of data—have proven to be an enormously effective way of recognizing objects in photos, identifying commands spoken into smartphones, and more. Now, researchers at places like Google, Facebook, and Microsoft are applying similar methods to language understanding, looking to identify patterns in English conversation, so far with limited success.
Mordatch and his collaborators, including OpenAI researcher and University of California, Berkeley professor Pieter Abbeel, question whether that approach can ever work, so they’re starting from a completely different place. “For agents to intelligently interact with humans, simply capturing the statistical patterns is insufficient,” their paper reads. “An agent possesses an understanding of language when it can use language (along with other tools such as non-verbal communication or physical acts) to accomplish goals in its environment.”
With early humans, language came from necessity. They learned to communicate because it helped them do other stuff, gave them an advantage over animals. These OpenAI researchers want to create the same dynamic for bots. In their virtual world, the bots not only learn their own language, they also use simple gestures and actions to communicate—pointing in particular direction, for instance, or actually guiding each other from place to place—much like babies do. That too is language, or at least a path to language.
Still, many AI researchers think the deep neural network approach, figuring out language through statistical patterns in data, will still work. “They’re essentially also capturing statistical patterns but in a simple, artificial environment,” says Richard Socher, an AI researcher at Salesforce, of the OpenAI team. “That’s fine to make progress in an interesting new domain, but the abstract claims a bit too much.”
Nonetheless, Mordatch’s project shows that analyzing vast amounts of data isn’t the only path. Systems can also learn through their own actions, and that may ultimately provide very different benefits. Other researchers at OpenAI teased much the same idea when they unveiled a much larger and more complex virtual world they call Universe. Among other things, Universe is a place where bots can learn to use common software applications, like a web browser. This too happens through a form of reinforcement learning, and for Ilya Sutskever, one of the founders of OpenAI, the arrangement is yet another path to language understanding. An AI can only browse the internet if it understands the natural way humans talk. Meanwhile, Microsoft is tackling language through other forms of reinforcement learning, and researchers at Stanford are exploring their own methods that involve collaboration between bots.
In the end, success will likely come from a combination of techniques, not just one. And Mordatch is proposing yet another technique—one where bots don’t just learn to chat. They learn to chat in a language of their own making. As humans have shown, that is a powerful idea.