Obsessed with technology?
Subscribe to the latest tech news as well as exciting promotions from us and our partners!
By subscribing, you indicate that you have read & understood the SPH's Privacy Policy and PDPA Statement.
News
News Categories

Google’s open-source TensorFlow deep learning software is going to get a lot more powerful

By Koh Wanzi - on 14 Apr 2016, 10:39am

Google’s open-source TensorFlow deep learning software is going to get a lot more powerful

Image Source: Google

Last November, Google open-sourced TensorFlow to the entire world. TensorFlow is a set of software libraries that runs on top of Google’s deep learning neural networks, and making it publically available for free would enable developers to take advantage of pieces of its code, and various neural networking models and algorithms, to create their own applications with image recognition, speech recognition, and language translation capabilities (among other things).

But Google didn’t actually open-source all of TensorFlow – Google runs TensorFlow across thousands of computer servers, but the version released to the public could only run on a single machine. That’s changing now, because Google has released an updated version of TensorFlow that will be able to run across hundreds of computers at a time.

If that seems like a huge boost to developers, it really is. Deep learning software like TensorFlow thrives on huge troves of data; in fact, it’s the only way for it to become smarter. TensorFlow requires large amounts of data – for instance photographs – to reliably identify recognizable patterns like faces. But this training takes time, and the ability to run the software over multiple machines simultaneously would allow users to train smarter systems a lot quicker.

Google TensorFlow

According to Rajat Monga, the technical lead on TensorFlow, the delay in releasing a multi-server version of the software was because of the difficulties in adapting the software for use outside of Google’s own highly-customized data centers.

In a post on the Google Research Blog, the company also highlighted that TensorFlow has been extremely well received. It is the most popular machine learning framework on GitHub, and was even the most forked project – when developers take a copy of the source code and begin to develop it independently – on the site in 2015, despite being released near the end of the year in November.

Projects that already use TensorFlow include a Pong-playing program, and a neural network that invents fake but seemingly legitimate Chinese characters.

Designer David Ha used TensorFlow to create a program that can invent fake but realistic-looking Chinese characters. (Image Source: David Ha)

While TensorFlow isn’t the only such service available, its beefed-up capabilities are still good news for all of us. We appear to be in the midst of a new period of openness and technology sharing, at least in the field of deep learning and server design.

For instance, Facebook also open-sourced its Big Sur hardware design last year. And in addition to open-sourcing TensorFlow, Google joined Facebook’s Open Compute Project in March and has announced a new family of cloud computing services which allows anyone to pay to use Google’s own machine learning technologies.

We appear to be arriving at a consensus – sharing is the way forward for everyone after all.

Source: Google Research Blog

Loading...