Today Facebook's AI Research division adds to that list with the open-sourcing of a series of machine learning -- or "deep learning" -- modules created for the Torch scientific computing framework.
According to a blog post provided by Facebook, those modules include "GPU-optimized modules for large convolutional nets (ConvNets), as well as networks with sparse activations that are commonly used in Natural Language Processing applications." Convolutional nets are used mainly in image- and video-recognition technologies; the latter networks are used to create algorithms that can train themselves on new data without needing an inordinate amount of human guidance.
Facebook claims the Torch modules it's open-sourcing are "significantly faster" than the ones bundled by default with Torch. The speed boosts come from GPU optimization and the ability to work across multiple GPUs in parallel -- and from using a Facebook-authored GPU algorithm that Facebook claims is much faster than the stock Nvidia GPU libraries for those processes. All of the work in question has been released through the company's GitHub account under the BSD license, so it's widely reusable with minimal restriction.
Facebook is banking on its investments in machine learning paying off in multiple categories. Aside from using ML for obvious moneymaking strategies like intelligent ad placement, the company wants to make a name for itself as an overall ML/AI pioneer and has already hired major-league talent to that end. Two professors from New York University, Yamn LeCuin and Rob Fergus, were brought on board at the end of 2013 to lead Facebook's AI research efforts in facilities in New York City; Menlo Park, Calif.; and London.
What's less clear is the long-term picture for Facebook's plans to push the envelope in AI and ML. In the most likely scenario, its innovations in those fields will be rolled back to its namesake social network, where the company's applications for the work will remain relatively closely guarded -- even while the algorithms and tools themselves are released as open source projects.
Contrast this with IBM's approach, where the innards of its Watson machine learning project are a closely guarded secret, save for some of its originating open source projects. But Watson itself can be consumed as a service, even if the roster of services provided under Watson is still limited. If Facebook offers a form of AI or ML as a service, it's likely to be tied strongly to its identity as a social-networking company -- such as external facial recognition APIs that leverage Facebook's data set -- than one along the lines of Watson's more open-ended business tools.