tensorflow.js: model training and predicting at the edge

VideoSlides


I did a talk on why Javascript is going to be important to the future of machine learning at Mesosphere last month. Thanks again to Chris Fregly for setting things up!

TL/DR: for data science, most people prefer python wrappers to simplify dealing with more complicated tools. Python has historically been a great wrapper language for dealing with Tensorflow and Javascript has been looked down upon as a toy language.

Basically, I agree that Javascript has a lot of faults, but on the other hand has a massive number of developers working on it daily. What this means in practice is that now that the team at Google has done the hard work of allowing us to write machine learning scripts in Javascript, all that remains is for people to start converting their workflows to the new language.

Under the hood, all these scripts generate machine code in some form or another. Time will tell, but things like keras.js show that intrinsically, there’s nothing that you can do in other programming languages that can’t be done in Javascript as well.

I’m not jumping programming languages quite yet, but long term it is my belief that the large number of developers that using Tensorflow.js will allow to start experimenting with machine learning will ultimately mean that soon Javascript could very well be the dominant language in this space. It’s worth your time to click around the tutorials at the very least!