A race is underway toward the future of computer technology with advances in a branch of artificial intelligence known as machine learning. Machine-learning software is trained to handle vast amounts of data, and then learns as it goes, often on its own.
Machine learning has been around for a long time, and it has been a crucial technology in the success of Internet giants like Google, Amazon and Facebook — used in the development of search, ad targeting and product recommendations. But in the last few years, machine learning has made huge improvements in computer vision, language translation and speech recognition, largely by applying the techniques of deep learning, which is inspired by theories about how the brain recognizes patterns.
Every major technology company is investing aggressively in artificial intelligence and machine learning. And not just computer companies. Last Friday, Toyota announced it would spend $1 billion for research and development on artificial intelligence in the United States over the next five years.
Google announced on Monday a bold step to establish its leadership in the field of machine learning, accelerate the pace of innovation in the field and potentially strengthen its business. It is making the software of its new machine-learning system, TensorFlow, which was developed over years, open-source code. The software will be freely available for outside programmers to use and modify.
Google announced the move on its official blog, noting that TensorFlow is “faster, smarter and more flexible than our old system” — up to five times faster in building and training machine-learning models. Its previous system, DistBelief, developed in 2011, was tailored for building neural networks, the building blocks of deep learning, and for use on Google’s own network of data centers. The new system, according to Google, is a far more general tool. “It may be useful,” the blog post said, “wherever researchers are trying to make sense of very complex data — everything from protein folding to crunching astronomy data.”
TensorFlow software, Google said, will be able to run on a single smartphone or across thousands of computers in data centers. The initial release of TensorFlow will be a version that runs on a single machine, and it will be put into effect for many computers in the months ahead, Google said.
There are other open-source software frameworks for deep-learning applications including
Theano, Caffe and Torch. But Christopher Manning, a computer scientist at Stanford University, who has tried TensorFlow, is impressed by the software. “It’s a better, faster set of tools for deep learning,” Mr. Manning said. “I think it will be extremely widely adopted by researchers and students in universities and in companies.”
In a separate post on Google’s Research blog, Rajat Monga, a software engineer, wrote, “TensorFlow is great for research, but it’s ready for use in real products too.”
With TensorFlow, Google is taking a page from its Android playbook. Android, the company’s mobile operating system, is open-source software and now the most widely used operating system in smartphones worldwide. And while Google does not make money directly on Android, the company profits handsomely from its search-advertising services on Android phones.
“The software itself is open source, but if this is successful, it will feed Google’s money-making machine,” said Michael A. Cusumano, a professor at the Massachusetts Institute of Technology’s Sloan School of Management. “There are so many applications of machine learning to the bread and butter of what Google does.”
Android is a technology used by millions of consumers, while TensorFlow is a technology for a smaller, if vital, marketplace — scientists and software developers. And each of the big technology companies is trying to woo software developers to its machine-learning technology. Microsoft has Azure Machine Learning, Amazon has Amazon Machine Learning, and IBM has Watson.
Google’s move, said Oren Etzioni, executive director of the Allen Institute for Artificial Intelligence, is “part of a platform play” to attract developers and new hires to its machine-learning technology. “But Google is taking a much less restrictive approach,” Mr. Etzioni said, “than just unveiling some services linked to your cloud offering, which is what IBM, Microsoft and Amazon have been doing.”
Computer scientists will be trying Google’s code and gauging its performance, but they will also be closely watching how the company manages TensorFlow as an open-source project. “This platform will live or die based on how they handle who controls updates to the code,” said Gary Bradski, a computer scientist who is president of OpenCV, an open-source project for computer-vision software. “Can the community have a say, or will Google control the official version by fiat?”