At a Machine Learning @Scale event today, Facebook revealed that it is now making use of artificial intelligence (AI) to identify photos on the social media network. While that may be of keen interest to marketers and people who spend time manually attaching tags to photos to make it simpler for others to find them, it more importantly suggests that open source AI technologies that Facebook shares with other companies are starting to mature.
Joaquin Quiñonero Candela, director of applied machine learning at Facebook, says that in addition to being able to identify photos, Facebook has a prototype of those same types of capabilities working in its labs that can be applied to video.
The core FB Learner Flow technology being employed by Facebook was created by Facebook AI Research, which shares its research and development work with other companies. FB Learner Flow makes use of an automatic alt text capability that Facebook developed to make it easier to identify photos without relying on tagging them. Facebook is also working on a Lumos project that applies deep learning technologies to provide more context about an image or video.
Facebook’s decision to start using these technologies in production is potentially significant because many organizations are trying to decide whether they should employ advanced AI services being made available via application programming interfaces (APIs) by vendors such as IBM, versus building out their own capabilities.
It may take a while for advanced AI technologies to be broadly applied, but Candela says Facebook tries to focus its AI efforts on narrow sets of tasks.
“Our advice is to focus on well-defined problems,” says Candela.
At the same time, however, Candela revealed that Facebook is already applying AI to 1.2 million models that are being trained to perform a specific function every month. Those instances involve everything from identifying photos on Facebook for the visually impaired to eventually identifying “fake news” content.
While IBM deserves a lot credit for advancing many of the core concepts, it’s becoming increasingly clear that open source projects are starting to close the gap in terms of core capabilities. Those open source projects come with the benefit of not being attached to a service that requires monthly payments to invoke. Regardless of the approach pursued, however, it might not make much sense to reinvent AI wheels that are already readily available.