The corporation describes TensorFlow as faster, smarter, and more flexible than its predecessor, capable of allowing the company to build and train neural networks up to five times faster compared to the first-generation system. It’s also just slick. Any computation that can be expressed as a computational flow graph can be computed with TensorFlow. Additionally, it provides tools to construct subgraphs common in neural networks, and will also compute derivatives for you once you’ve defined the computational architecture and added the necessary data.
Under the hood, it features a production-grade C++ backend that runs on Intel CPUs, Nvidia GPUs, Android, iOS, and OSX. As for the Python front-end, TensorFlow interfaces neatly with Numpy, iPython Notebooks, and an armada of related scientific tooling. Google’s Jeff Dean recently told Wired that the company will provide sample neural networking models as well, including “models for recognising photographs, identifying handwritten texts, and analysing text.”
It should be noted that the company isn’t open-sourcing everything. Wired writes that the current version only runs on a single computer, although that might change in the future. Google clearly wants the machine learning community to help build TensorFlow into a more mature tool that can accelerate certain fields of research and development. Oren Etzioni, executive director for the Allen Institute for Artificial Intelligence, speculates that it’s “part of a platform play” intended to attract developers and new hires.
