One way I could see this being cool is to drastically cut down the training time of some specific task. DeepMind's model can do a lot of things, but none particularly well. It would be nice if you could start with their model weights and update them to perform well on your specific task with your data. Ideally this process would be cheaper than starting the training from scratch. I also think this is how the human mind develops. There is some intrinsic knowledge that a human starts with, e.g. how to recognize a face and how to grasp, but then training happens over the course of one's life.
No comments yet.