About
This sample uses a deep neural network to predict the action you will perform based on the history of previous actions you've performed. It records every page you visit and uses that information to continue learning and improve its predictions. Learning is performed asynchronously while the browser is idle so performance is not affected. The entire neural network is constructed on the browser and is WebGL accelerated for fast performance. No backend server is needed.
Usage
Choose a few patterns and click around the site. Your actions will be recorded to local storage. Once you have about 100 clicks, hit the "Learn" button. This will create and train the initial model. After that, when you click around, a graph of predicted next actions will be shown, and "Quick Actions" will be displayed in the lower right corner.
Predictions
Training
Once the model is initially trained, it will automatically update itself as your history is updated.
To train the model from scratch, press the button below. Results are recorded in the console. In general, you should have at least 100 interactions before the model will be very accurate.