TensorFlow: machine learning from Google, now smarter for everyone

Just a couple of years ago, we couldn’t communicate with Google applications through the noise of the street, didn’t translate Russian inscriptions on Google Translate and didn’t look for photos of the same labradoodle in Google Photos, just hearing about it. The fact is that our applications were not smart enough then. Well, very quickly they became significantly, much smarter. Today, thanks to machine learning technology, all these wonderful things, as well as many other and more serious things, we can do effortlessly.

In general, welcome: we have created a completely new machine learning system called TensorFlow . TensorFlow is faster, smarter, and more flexible than our previous technology ( DistBelief , since 2011, the one thatrecognized a cat without a teacher ), making it much easier to adapt it to use in new products and research projects. TensorFlow is a highly scalable machine learning system that can work both on a simple smartphone and on thousands of nodes in data centers. We use TensorFlow for the whole spectrum of our tasks, from speech recognition to an answering machine in Inbox and search in Google Photos. Such flexibility allows us to design and train neural networks up to 5 times faster in comparison with our old platform, so that we can really use the new technology much faster.




Using the new “straight from the workshop” system, we saw what opportunities TensorFlow offers. It became clear that she was able to exert even greater influence outside our corporation. Today we open TensorFlow for all developers in open source. We hope that this will allow the entire community around machine learning, everyone - from researchers to engineers and just amateurs - to exchange ideas more efficiently, through really working code, and not through research articles. As a return on this step, we plan to use your feedback to speed up our research in machine learning, ultimately making this technology better for everyone. By the way, a bonus: TensorFlow is not only suitable for machine learning. It can also be useful in all cases when researchers look for content and meaning in very complexly structured data - from protein bundles in bioinformatics to analysis of astronomical tables. While the previous technology was focused on neural networks and our internal kitchen, TensorFlow is quite generalized and more efficient.

Machine learning is still new and very young. Today, the computer is still not able to repeat what a 4-year-old child does without effort. What is it, for example, to remember the name of a dinosaur after seeing only a couple of presented pictures, or to understand that “The mass of the working glass” is not at all about the working man, but “These types of steel are in stock” is not about eating metal, and most likely not about people who came to the warehouse to eat. Ahead is a lot of work in this direction. We believe that TensorFlow provides a wonderful starting point at the beginning of the journey, as well as the opportunity for us all to walk close roads in the right direction. Join now!