Google has decided to release to the open-source community the machine learning system at the core of many of its services, including Google Translate, Smart Reply in Inbox and image search in Google Photos.
Google's second-generation machine learning technology, dubbed TensorFlow, is designed to run on everything from a single smartphone to thousands of data center systems, CEO Sundar Pichai said in a Nov. 9 blog post.
TensorFlow is a faster, smarter version of Google's first-generation machine learning technology and allows the company to build and train neural nets at five times the speed than what was possible previously, Pichai said. The technology is also flexible enough to be adapted to new products and applications, he said.
By releasing TensorFlow to the open-source community, Google wants to make the technology available to engineers, hobbyists, academic researchers and others in the machine-learning field. The goal is to enable the exchange of ideas "much more quickly, through working code rather than just research papers," Pichai said "And that, in turn, will accelerate research on machine learning, in the end making technology work better for everyone."
TensorFlow can be used in applications other than machine learning. Any application that requires researchers to make sense of extremely complex data sets, like protein folding and crunching astronomical data, can benefit from TensorFlow as well, Google's CEO added.
The company's first-generation deep learning infrastructure, dubbed DistBelief, was developed in 2011. Google has used the technology to build massive neural networks designed to process information in a manner similar to the human brain. Google has used DistBelief to show, among other things, how a neural network was able to teach itself to recognize images of cats from still frames in unlabeled YouTube videos.
Google claims that DistBelief helped the company improve its speech-recognition technology by 25 percent and to implement an image search capability in Google Photos.
Despite its enormous success, DistBelief had its limitations, Google Senior Fellow Jeff Dean and Technical Lead Rajat Monga said in a blog post, also published Nov. 9. "It was narrowly targeted to neural networks; it was difficult to configure; and it was tightly coupled to Google's internal infrastructure—making it nearly impossible to share research code externally," the two Google researchers said.
The open-source TensorFlow addresses such limitations while also delivering better speed and scalability, they noted. In fact, in some benchmarks, TensorFlow turned out to be twice as fast as DistBelief.
According to Google, the machine learning system, as released to the open-source community, is ready for use in real production environments. The company has released sample model architectures that enterprises can use to get started quickly with TensorFlow.
Google's decision to open-source TensorFlow could accelerate development around machine learning. But it is not the first company to do so. Earlier this year, Facebook's AI Research (FAIR) team open-sourced some of the deep learning modules that it had developed around computer vision, machine learning and numerical computation. At the time, Facebook had noted that its decision was motivated by a desire to stimulate progress in the deep learning field.