Except, it doesn't render anything. It's the recently used parts of the higher layers of the inference network starting to have electrical activity of their own to do some kind of optimization/defragmentation.