Daedalus Generating Mazes With Autoencoders and Variational Autoencoders
Daedalus, master craftsman of ancient myths, conceived the Labyrinth: a maze of bewildering complexity. Its winding paths and endless turns, a testament to his genius, were designed to confine the fearsome Minotaur, blurring the line between architectural marvel and cunning trap.
An autoencoder is a type of network shaped like an hourglass. We start with an input, and pass it through smaller and smaller layers until reaching the bottleneck point (also refered to as latent space). We then feed the bottleneck to bigger and bigger layer until we reach the original size. And train the network to recronstruct the original input.
By tweaking the values of the latent space, the hope is that our second half of the network (the decoder) would re-create a valid output, but somewhat different from the training set.
We’ll take a detour through MNIST because it’s a canonical example and I want to have a way to evaluate how poorly is the network doing.
