The simulation below is a 1-dimensional regression where a neural network is trained to regress to y coordinates for every given point x through an L2 loss. That is, the minimized cost function computes the squared difference between the predicted y-coordinate and the "correct" y coordinate. Every 10th of a second, all points are fed to the network multiple times through the trainer class to train the network.
The simulation below will eval() whatever you have in the text area and reload. Feel free to explore and use ConvNetJS to instantiate your own network!
Add data points by clicking!