Last week, Facebook unveiled its virtual assistant on a few hundred phones in the San Francisco Bay Area. What makes its AI, called "M," different from the likes of Siri? It hasn't eliminated the human element.
Other AI, like IBM's Watson, attempt to think how humans think with little human intervention. Facebook's M, which is built on its Messenger app, works "in tandem" with a team of humans. Facebook acquired startup Wit.ai to build this tool.
“The AI tries to do everything,” says Alex Lebrun, the founder of Wit.ai, told Wired. “But the AI is supervised by the people.”
While M will be capable of answering most questions posed to it and do a wide variety of tasks, there are gaps in its web of knowledge, and that's where its team of human beings comes in. When a user asks M a question, the software sends the question and a proposed answer to a member of the team who then checks to see what else may be needed to fully answer the query. The team of flesh and blood people working on the assistant is training and improving its (Her? His?) knowledge with each unknown piece of data by giving the information to it directly.
Rather than introducing fancy new technology, M works to perfect what's already been tried and tested. Lebrun started Wit.ai in 2013, first building artificial intelligence that helped customers of companies like AT&T with their needs via speech recognition, including natural language. The software proved to be more lightweight because it didn't rely on vaults of data.
M uses conditional random fields and maximum entropy classifiers as part of the core of its technology. By joining together with humans to reach the correct answers, it develops a neural net to learn how to search for the right answers for future questions of a similar nature. Such a system is described as deep learning, or a "form of AI that masters tasks by analyzing enormous quantities of information across a vast network of machines."
“This is a good way to bootstrap. With a few thousand data-points, you can start to build a model,” Lebrun said. “Then, using this model, you get more data, and once you have about a million data points, you go to [Facebook computer scientist] Yann and get some deep learning.”
To read the full Wired.com article, click here.