Facebook is open-sourcing its AI bot-building research
The fastText library can cut language training from days to seconds
Say hello to smarter artificial intelligence. Soon, anyway. Facebook is opening up the secret sauce that powers its bots so the public can employ and study it. This is part of the Facebook AI Research (FAIR) lab's mission to help researchers and engineers by making its work available to all. Called fastText, the library of code is now available on Github for public use and scrutiny, and will require a compiler with "good C++11 support."
The biggest appeal of fastText appears to be speed and efficiency. According to Facebook, fastText is, as its name suggests, much quicker than other learning methods, and can train models "on more than 1 billion words in less than 10 minutes using a standard multicore CPU." In fact, FAIR claims that, compared to deep learning models, fastText can cut training delays from several days to a few seconds.
fastText focuses on classifying words and sentences, and produces libraries that programs can reference when executing tasks. For example, fastText can learn that the words "boy," "girl," "man" and "woman" refer to specific gendered nouns and store those values in a document. Then, when an AI program, like a bot, is interpreting a request, such as "Where my girls at," it can look into the fastText-generated document and understand that the user is asking for female names.
It's easy to see how this move makes sense for the social network. It started integrating chat bots in its Messenger app this year, and making it easier to train AI can boost the growth of third-party offerings. Considering the proliferation of AI integration in many of its competitors, such a move can also encourage developers to focus on building for Facebook's platform first.
In a statement, FAIR said, "Ultimately, we hope that fastText will help us all design better applications and further advance the research in language understanding." Perhaps future AI developers can look to FAIR's research for help, which for now appears to be a far more sensible resource than Reddit.
