An Introduction to TensorFlow.js with Nick Kreeger.

Sept. 3, 2018, 2:04 p.m. By: Kirti Bakshi


In this video of TensorFlow.js talks at SFHTML5, the two speakers Nick Kreeger (@nkreeger) and Ping Yu (@pyu10055) come forward and not only share their TensorFlow.js expertise but also answer various questions and therefore, present a very fine introduction to Tensorflow.js.

Introduction to TensorFlow.js with Nick Kreeger:

Talk #1—Introduction to TensorFlow.js with Nick Kreeger:

This talk will dive into the origins of TensorFlow.js and what the platform delivers to all the developers worldwide today. Runtimes for browser and server side will be highlighted as well with the help of demos, concepts, and sample APIs.

About the speaker:

One of the speakers in this video, Nick Kreeger (@nkreeger) is a Senior Engineer at Google Brain, TL for TensorFlow.js - previously TL on the Google Store, Strava, Rdio, and Songbird (Mozilla).

Now let's move onto talk #2:

Talk #2—Audio Model with TensorFlow.js:

TensorFlow.js that is the recently-released JavaScript version of TensorFlow runs in the browser and Node.js. In this talk, the main aim is to not only introduce the TensorFlow.js Machine Learning framework but to also show how to actually perform the complete Machine-Learning workflow, including the training, client-side deployment, as well as transfer learning.

The focus was kept on the machine-learning task of word recognition making the use of audio inputs, utilizing the Speech Command Dataset (Warden, 2017). A Convolutional neural network (CNN), is defined by making the use of the Keras-like high-level API of TensorFlow.js.

The input to the CNN are the spectrograms that are generated from the web browser’s WebAudio API. The training process takes place in Node.js, which directly binds to the CUDA-accelerated kernels of TensorFlow. The trained model is then converted into a browser-friendly format and deployed in the web browser for low-latency, real-time on-device inference accelerated by WebGL.

In addition, you are also shown that the deployed model can be fine-tuned and saved at the client side, therefore, achieving recognition of words outside the original vocabulary.

All appreciation goes to the cross-platform nature of the browser, as both the deployment and fine-tuning can occur in supported web browsers on a wide range of devices.

About the Speaker

The other speaker in this video, Ping Yu (@pyu10055) is a Senior Engineer at Google Brain, a core team member of TensorFlow.js, and was previously the TL for Google Attribution.

Video Source: SFHTML5