Unique to transfer learning is the concern about how similar your source and target problem domains are to one another. A classifier trained to recognize faces probably won't transfer easily to a target domain recognizing various architectures. We ran experiments where the source and target were as different as possible, as well as experiments where the source and target domain were very similar. Unsurprisingly, when the source and target domains in the transfer learning application are very different they require more data than when they are similar. They also require much more fine-tuning, since the feature extraction layers have a lot of relearning to do, when the domains are visually very different.