Scrap Mechanic | Recurrent Neural Network Concept

Technically speaking it is not a neural network, because learning mechanism is not implemented. But instead I added feedback link, which is responsible for some of learning mechanics.

Each neuron consists of 20 bearings and 20 controllers. More than 3 neurons causes massive framerate drop. That's why there is only 3 of them. 1 neuron recognizes 1 character/pattern, so this machine can recognize only 3 characters at a time. But if there was no lag, it could be easily expanded.

Since there is no learning mechanism (except for setting weights manually), I added a feedback layer consisting of 1 neuron. Small neuron in the feedback layer activates only if there is more than one activated neuron in the first layer. If so it sends to neuron inputs in the first layer a signal, which gradually decreases weighted sum of each neuron. It will continue until only one activated neuron in the first layer is left.