This also applies to the styles acquired by neural networks: given some training data and a network architecture, there are multiple sets of weights values (various models) which could clarify the information, and easier models are more unlikely to overfit than advanced ones. In the above mentioned display grab, we can see the memory progressively