Subscribe to download
Perceptron neural network

This level uses a neural network based on a single-layer perceptron to classify numbers given as an input.

To use it, build numbers.

When you start the sim it will try to guess what the number is. Note however that this neural network is quite simple (I only figured out how to make one layer efficiently), so the guessing will be quite bad. Also note that the position of the number will matter. The network was trained on the "MNIST" dataset, which is a dataset that has many hand-written numbers.


So try to build numbers that resemble them for the best results.

The accuracy of the network on the MNIST test data was 81.74% which decently high, but it isn't perfect, meaning that it will guess incorrectly. Even when numbers are quite similar to the MNIST images. Level logic also appears to be limited to two decimal points of precision, meaning that the actual level neural network will perform even worse.

IMPORTANT: All the neural network logic that is used in the level is inside a block underneath the map. If you edit this block the game *will* freeze.