This is one of the greatest challenges posed. There really is no determinant in what makes the current network sufficient or similar enough to the one you are trying to apply. In fact, the algorithm can perform even worse if it fails to find this connection. The model may even do worse than if it was never even trained in the first place.