Artificial neural variability for deep learning: on overfitting, noise memorization, and catastrophic forgetting