Leveraging TensorFlow.js to improve User Experience

Oliver Zeigermann

In typical offline scenarios with a fixed set of data, machine learning on powerful servers is the way to go. Yet, when dealing with interactive Single Page Applications, zero installation, low latency and high privacy concerns change the story towards machine learning in the browser with TensorFlow.js.

In this talk I’ll show how to train RNN, LSTM, and GRU based models both on client and on server using mouse movements to infer which button the user is going to click. We will discuss potential applications, including highlighting of buttons for easier access and help with complex flows.

As training data will be different for each user, special attention will be put on model evaluation.

Objective of the talk

To look into when machine in the browser does make sense, finding out how it works, and what is needed for it to succeed. We’ll also talk about training models with small data sets and making the results robust and explore how different RNN variations impact the model for sequential data.

Required audience experience

Basic knowledge of machine learning and neural networks.



Track 3
Location: Burton and Redgrave Date: October 1, 2019 Time: 2:30 pm - 3:15 pm Oliver Zeigermann Oliver Zeigermann