TensorFlow recently open sourced TensorFlow.js. This is an open source library that helps us define and train deep learning models entirely on the browser using JavaScript as well as through a high-level layered API. We can use this to train deep learning models entirely on the client side. A server GPU is not required to train these models.
The following diagram illustrates the API overview of TensorFlow.js:

This is powered by WebGL. It also provides a high-level layered API. It has a support for Eager execution, too. You can achieve three things using TensorFlow.js:
- Load existing TensorFlow/Keras models for in browser predictions
- Retrain existing models using client data
- Define and train Deep Learning models from scratch on the Browser