oh my darwin @ylecun, you should watch what you said in our debate, in reply to Chalmer’s query about my list of nativist claims: http://tinyurl.com/lecunmarcusdebate …, perhaps 1:15 in, where you denied all my nativist proposals, even convolution. please reconcile with what you say below.https://twitter.com/ylecun/status/957766752481443845 …
-
The debate was about how much prior structure you need, and what type, not whether you need it. But the answer depends on how much training data you allow yourself to use. The more data you use, the less structure you need. Even convolutional structure can be learned.
Has anyone learned convolutional structure in a way that generalizes outside the space of training examples? Interested in references.