This is a transcript of a conversation with deep learning pioneer Geoffrey Hinton and I have to say, it's riveting reading. It begins by discussing capsules: "what capsules are trying to do is recognize whole objects by recognizing their parts and the relationships between the parts." Google recently filed a patent for this idea. Once you have a set of capsules you're confident belong to a single object, you begin assigning names to them (when we're learning a language, we're not learning what things there are, we're learning what names to give things we already know exist). That's just the tip; there's a great section in the middle about the (non-)role of back propagation in human learning and Hinton's own (unsuccessful) theory of perceptual learning. And much much more.
Today: 5 Total: 1101 [Direct link] [Share]
