I just say that there's no introspection in existing networks. There's no part, which can take the weights of a convolutional part of a network and images of a

I just say that there's no introspection in existing networks. There's no part, which can take the weights of a convolutional part of a network and images of a certain class, and then output confidence score of the convolutional part classification ability for that class. I don't say that it is possible or useful. I said that I don't know of any deep learning systems, which "[...] can freely introspect their thought process as part of their thought process."