Google (the real Google, withinAlphabet) has announced a new, open-source machine learning platform called TensorFlow, one which could greatly accelerate the pace at which neural networks are taking over the search giant’s service portfolio — and its business model. The company claims it can significantly improve both the speed and power of neural network development across a wide variety of platforms and applications. If Google and the community can develop the platform as they clearly hope, then TensorFlow could become Google’s new organizing principle, and a major driver in the continuing evolution of the Internet.
Google is calling it the company’s “second generation” machine learning platform, successor to the successful but now-aging DistBelief platform that has led to many of the company’s current services. It says that by using TensorFlow, its developers can build and train a machine learning algorithms “five times faster” than previously possible.
That’s an important advance, since Google’s machine learning initiatives are some of their most important, at this point. Machine learning is increasingly how Google sifts the mountains of data we provide for them, how it pulls salable signals out of seemingly endless volumes of noise. Machine learning lets the company burrow ever-more-invasively into people’s lives by providing services too interesting and valuable to pass up, from translation to facial recognition. It’s also the main technology driving the epic Now-versus-Siri-versus-Cortana triforce of corporate one-upmanship, which could very well end up determining many users’ choice of mobile platforms over the next five years.
These solutions are coming to define not just the services Google provides, but the methods by which it provides them and coordinates their findings. TensorFlow could very quickly become Google’s new brain, and by extension a meaningful upgrade to the Internet overall.
So, what the hell is TensorFlow? TensorFlow is a library of pre-built portions of neural network code with easy-to-use tools to customize them deeply, and add to them with as much flexibility as possible. This is not really a new idea in the context of other open-source machine learning platforms, like Torch — but this is Google, and as such it’s not unreasonable to assume that its standard will become the standard. And since Google can easily attract an army of eager, talented coders from the open source community, it seems the most likely to progress the most quickly. They’ve designed TensorFlow to accept two of the most widely used programming languages, Python and C++.
TensorFlow scales to run on everything from desktop super-crunchers to laptops to smartphones. The program lets developers use their tablet to fiddle with a program’s design on the bus, then switch seamlessly to running or training that algorithm on a much more powerful desktop when they get home.
TensorFlow also divorces Google’s machine learning workflow from its monolithic company codebase, meaning that it’s now possible for outsiders to meaningfully contribute to the project. DistBelief was not user-friendly, and its less forward-thinking design made it “nearly impossible” for the company to share its research code externally. With TensorFlow, we could hypothetically see collaboration with the community lead to an explosion of sophistication in machine learning. And knowing Google’s approach to hiring, such an important contribution to the company might function as a new space for talented coders to distinguish themselves in the company’s eyes.