http://people.ciirc.cvut.cz/~hlavac/pub/MiscTextForStudents/1982MarrDavidVisionBook.pdf
Seymour Paper wrote the summer project computer vision memo: https://people.csail.mit.edu/brooks/idocs/AIM-100.pdf
pg 12, frogs respond to black disks which represent bugs. Barlow’s 1953 study of ganglion cells in the frog’s retina. Ganglion cells output signals from the retina to the brain. It is like these cells do first stage object detection, like region/candidate generation. The brain would do the more finegrained processing.
It was, of course, recognized that physiologists had been lucky: If one probes around in a conventional electronic computer and records the behavior of single elements within it, one is unlikely to be able to discern what a given element is doing. But the brain, thanks to Barlow's first dogma, seemed to be built along more accommodating lines-people were able to determine the functions of single elements of the brain.
The problem of understanding how a neuron does something is a shared problem between neuroscience and deep learning. We can identify a neuron that responds to a hand but not understand how its does that as easily.
Three levels
Human vision is general but still have some specialized function. Our attention quickly goes to objects when we sense sudden motion.
The brain stores information about the geometry of an object separately from information about its use or semantics (35). Semantic knowledge can be inferred when there is limited geometric knowledge.