Tommy S

"Did you read the Nature Neuroscience article, or the pop sci summary? I can give you a pdf of the publication if interested. As far as I can tell, the training & "test" data are independent - the network isn't trained at all to identify numerosity, it's trained to reconstruct the visual input. And that's it. Number of objects plays into that learning rule zero (especially if surface area, density, size, etc are independent of numerosity, which is controlled for). Even as such, there are nodes of the network which best describe the numerosity of the stimulus, independent of those properties, and they were not trained to have this property at all. Full disclosure, I don't work w/ neural networks daily or anything, but I'm familiar w/ how they operate and the importance of testing a trained network w/ novel data. This looks legit, and is a cool and important new finding." - Tommy S

"submitted by reddit to blog." - Tommy S