Also, got some progress on my side project.

The cnn based auto encoder works better without the dense layer (or full connect layer), I don't know why, but it just works.

Here are some examples, the green pixel means the autoencoder thought there should be a living cell, and it is. Red pixel means the autoencoder think it's alive, but it's not. White means the autoencoder missed, it though it's dead, but it's alive.

Next step: try to build a neural network that can give necessary addition to make next step as close as possible to a given target pattern. (So I can display the bad apple in Conway's game of life). And to make things less boring, I decide to use genetic algorithm to "train" the network. Wish me luck 🤣