Not too surprising, actually. Take images. Once you have a deep convolutional neural net trained on one set of image data, switching to another set should be fairly straightforward as I’d think a large portion of the existing net has learned edge, shape, feature, and object detection.

Now it’s more of a matter of cross-matching shapes and features to a new label set.

Makes me wonder what would happen if you took the aforementioned trained model, re-randomized the last dozen or so layers, and then started over with a new data set.