Google gives everyone machine learning superpowers with TensorFlow 1.0

This site may earn affiliate commissions from the links on this page. Terms of use.

Information technology wasn't that long ago that building and training neural networks was strictly for seasoned computer scientists and grad students. That began to change with the release of a number of open-source car learning frameworks like Theano, Spark ML, Microsoft'south CNTK, and Google'southward TensorFlow. Among them, TensorFlow stands out for its powerful, yet accessible, functionality, coupled with the stunning growth of its user base. With this calendar week'due south release of TensorFlow 1.0, Google has pushed the frontiers of machine learning further in a number of directions.

TensorFlow isn't just for neural networks anymore

In an effort to make TensorFlow a more than-general motorcar learning framework, Google has added both built-in Estimator functionality, and support for a number of more traditional car learning algorithms including K-means, SVM (Support Vector Machines), and Random Forest. While there are certainly other frameworks like SparkML that support those tools, having a solution that tin combine them with neural networks makes TensorFlow a great option for hybrid problems.

TensorFlow 1.0 also offers impressive performance improvements and scaling. In i benchmark, a training session running on a 64-processor machine ran nearly 60 times as fast as i running on a single processor.

With Keras, anyone has a adventure to build the side by side HAL9000

This is all the code needed to build a model that analyzes videos and answers questions As powerful as TensorFlow is, constructing a complex model straight in its API takes quite a bit of knowledge, and some conscientious programming. This is especially true of sophisticated models like recurrent neural networks and their fancy cousins, LSTMs (Long Short Term Memory models). The Keras programming interface provides a more than user-friendly layer on top of TensorFlow (and Theano) that make constructing high-end networks deceptively simple.

During the Superlative, Keras writer Francois Chollet showed how like shooting fish in a barrel it is to build a network that looks at video sequences and answered questions nigh them — in a single folio of code! Of course, knowing how to put various layers in the model together still takes a lot of skill, merely actually amalgam it is relatively painless. Keras also includes a number of pre-trained models for easy instantiation. Given the labor-intensive nature of assembling the large datasets needed to train models, and the processor-intensive nature of preparation, that's a huge benefit for developers.

Making your smartphone a lot smarter

One of the most impressive new capabilities of TensorFlow is that its models tin be run on many smartphones. TF1.0 even takes reward of the Hexagon DSP that is built into Qualcomm'south Snapdragon 820 CPU. Google is already using this to power applications like Interpret and Word Lens fifty-fifty when your phone is completely offline. Before now, sophisticated algorithms similar those required for translation or speech recognition required existent-time admission to the cloud and its compute servers.

TensorFlow has also been ported to IBM'south POWER compages as office of PowerAI, and to Movidius's Myriad 2 specialized processor.

Getting started with TensorFlow

You can download TensorFlow 1.0 at present. Currently, Keras is a split packet that'due south easy to install using pip or your favorite package manager, but Google plans to take it born to the one.ii release of TensorFlow. There are some API-breaking changes going from .12 to 1.0, only many of them are fairly straightforward proper name changes that take already been telegraphed with Deprecated messages. Google even provides a handy script that will try and update your existing code, if needed.

As is typical of machine learning tools, yous'll get much better functioning running on a supported GPU, merely now there are even options to spin your models up in the deject. For example, Y Combinator-backed startup Floyd Hub has TensorFlow and many other car learning tools pre-installed on powerful GPU systems yous can rent just for the amount of time you need to train and run your models.

Now read: Artificial neural networks are changing the globe. What are they?