Basically your problem was your ReLU function. ReLUs suffer from "Dying ReLU" which basically is when the function sets the value to 0 which can cause neurons to "die". The fix was switching to "Leaky ReLU" which instead of setting to 0, it sets it to a small but positive value (0.01). Credit isn't needed, it's a simple architecture fix. I also added an output sampler inside, which is the average of the output vector list displayed in the "AI" variable.