The way I imagine it works is like the weights of a machine learning network. The lower-level convolutional layers are discarded and only the higher-level featu

The way I imagine it works is like the weights of a machine learning network. The lower-level convolutional layers are discarded and only the higher-level feature extracted weights (latent space I believe it's called) are remembered. We can then 'think' in that feature space with no associated imagery but feel our way around in that space. There was one time (before the web) where I was trying to solve a physical puzzle of balancing many long construction nails on the head of a standing one. I couldn't solve it, but must have thought about it because when I woke up, I knew the solution and demonstrated it (I already knew it would work) as soon as I got to the office where the nails were. I also spend long periods pondering a problem staring into space and not thinking of anything specific but during that time my mind is mixing and matching and trying different arrangements of possible solutions. I can't see any of this happening, but every now and then I'll have an idea making connections I hadn't considered before. Often one of them will pan out well.