Google releases TensorFlow.js to bring machine learning to the browser

The 2018 TensorFlow Developer Summit was held on March 31st at the Computer History Museum in California, USA, and brought together global machine learning developers for a one-day technical sharing and demonstration.

Google Announces TensorFlow.js, a Machine Learning Framework for JavaScript Developers 

It’s easier to do machine learning on your browser! Google released TensorFlow.js, TensorFlow technology combined with Javascript, significantly reducing the threshold for developers to develop machine learning in the browser. Google stated that with the development of Javascript and machine learning technology, all the work of machine learning can already be fully implemented on the browser, including defining, training, and running machine learning models.

Google engineers Nikhil Thorat and Daniel Smilkov, who live broadcast at the TensorFlow Developers Summit, used TensorFlow.js with camera and computer vision technology to teach artificial intelligence programs to play PAC-MAN games in a full browser environment. Sample programs have been open sourced on Github.

Google stated that running machine learning in the browser means that there is no need to install any libraries or drivers. Just open the web page and the program will run. In addition, TensorFlow.js supports WebGL, so it can also use GPU-accelerated operations.

TensorFlow.js provides three workflows to handle the various stages of machine learning models. First, developers can convert TensorFlow or Keras pre-trained models into the TensorFlow.js format and load them into the browser for extrapolation. Second, the developer can not only load an existing machine learning model, but also use the image data collected from the user’s browser to train the model. This technique is called Image Retraining, Google says, PAC-MAN The game sample program is performed in this mode. The advantage of this method is that a small amount of data can be used to make the model more accurate.

Of course, the user can also build a machine learning model from scratch in the browser and use the API provided by TensorFlow.js to completely define, train and run the model in the browser. These APIs are very similar to those provided by Keras. The developers should be able to get started quickly.

TensorFlow for Swift will open source in April

Although this project is still in its infancy, many people are expecting it. For example, Jeremy Howard, the founder of fast.ai and the former Kaggle president, put this as the most important release of the summit, but also said: Can we finally put Python down?

There is less information about TensorFlow for Swift, and interested parties can visit this site.