6DNet is an educational game which is useful for understanding how neural networks and backpropogation work. In order for a neural network to minimize the difference between the expected output and the actual output, it can adjust the weights (represented by the sliders in the game) to achieve the lowest possible error. As you’ll find, some methods of combining the weights allow for lower error than others.

This specific network models the exclusive-or, also known as the “XOR”, network. The two inputs can either be 1 or 0, meaning there are four possible combinations. To achieve a perfect output (in this case, a 1) from the network, only one of the inputs can be a 1. 

Instructions:

Use the colors on each plane or the colored arrows alongside each slider to navigate the weights towards values which give a lower error. You can select a different weighting method to see the minimum error for each method. If you’d like to automatically solve the network, use the “auto solve” function.