They argue that model architecture and training procedure (algorithm) should play some role in generalization. Without these factors, current generalization results can become vacuous.

