Machine Learning and the Platonic Ideal of Rocky

How does the neural net in our skull grok something as complicated as “Rocky”?

WAUWINET, AUGUST 14, 2026.  I’m in a hospital room looking at a 21-year-old man looking at a laptop flashing images of action heroes, including the cinematic boxer “Rocky.” Cables loop from the back of the man’s bandaged head to a stack of blinking electronic equipment.

This is one of many strange, loopy memories stirred up by my strange, loopy attempt to learn how machines learn. I call this project strange and loopy because, well, think about it: The neural network in my skull is trying to learn how artificial neural networks learn. A strange loop indeed!

In Why Machines Learn: The Elegant Math Behind AI, Anil Ananthaswamy details how brain research has inspired neural-network research. Example: In experiments on animals more than a half century ago, neuroscientists discovered brain cells that detect straight edges. Like the edge of a table, door, picture frame, wall, building, road, tree trunk. Like a horizon.

This revelation led to a view of perception as a process of generalization, or abstraction. The most elemental edge-detection neurons detect only vertical edges. Or only horizontal edges. Or only edges oriented at an intermediate angle. And so on.

Higher-level neurons fire--meaning they emit an electrochemical pulse, or signal—when they receive signals from any edge-detection cells. These higher-level cells are “invariant to translation,” Ananthaswamy says, meaning the cells respond to edges regardless of orientation.

The invariant-to-translation cells could be said to get the idea or concept of an edge. Ananthaswamy notes that this hierarchical process—simple cells feeding raw data to invariant-to-translation cells--can in principle result in cells that recognize triangles, squares and more complex forms. Like, say, your grandma’s face.

“Grandmother cells” were posited in the 1960s as a joke, a reductio ad absurdum, an example of how brains definitely don’t work. Ananthaswamy calls grandmother cells “far-fetched neuroscience lore.”

It is this derogatory description that makes me remember the young man in the hospital room. I met him in 2005. His name was Danny. He was a 21-year-old college student being treated for epilepsy in Los Angeles. Research on Danny and other patients suggests that our brains do in fact contain cells that recognize specific people.

Neurosurgeon Itzhak Fried carried out this research on epileptic patients. Prior to operating on the patients, Fried implanted electrodes in their brains to pinpoint the source of seizures. With the patients’ consent, Fried used the electrodes—some of which could detect signals from single brain cells--to investigate how brains process visual information.

Fried’s team showed patients images of all sorts of things: chairs, numbers, roses, tigers and humans. The research identified individual cells attuned to a particular relative: sibling, father, mother and, yes, grandmother. Researchers also found self-recognition cells, which fire when a patient sees an image of herself.

Talk about loopy!

Other cells fired exclusively in response to a specific famous person: Bill Clinton, Arnold Schwartzenegger, Jennifer Aniston, Brad Pitt. A Bill Clinton cell, say, would respond not only to different photographs of Clinton but also to drawings and the spelled-out name: BILL CLINTON.

Talk about invariant to translation!

One of Danny’s cells fired only when he looked at images of Sylvester Stallone playing Rocky. (Danny loved Rocky.) Imagine the enormous hierarchy of invariant-to-translation cells required to produce a cell that encodes the concept of an imaginary boxer!

Back in 2005, I asked two of Fried’s collaborators, Rodrigo Quian Quiroga and my source and friend Christof Koch, to spell out the implications of what Fried calls “thinking cells.” Here is how I wrote up their responses in my 2005 article for Discover Magazine:

What excites Koch most about the thinking-cell results is the possibility that they may illuminate a fundamental component of cognition. Our comprehension of the world, he says, requires that we ignore much of the data flooding in through our senses. When we turn on a TV or reminisce about a movie, our brains somehow instantly compress raw sensory data into meaningful concepts and categories. This feat may be accomplished at least in part, Koch says, by cells that represent not just this or that particular image of Rocky but “the platonic ideal of Rocky.”

Quiroga notes that a short story by a fellow Argentine, Jorge Luis Borges, spelled out what would happen to us if we lacked this capacity for compression. “Funes, the Memorious” tells the tale of a youth who, after falling from a horse and striking his head, becomes gifted, or cursed, with photographic recall of every minute experience. He is so overwhelmed by the infinitude of his perceptions that he retreats into a darkened room. “To think is to forget a difference, to generalize, to abstract,” Borges writes. “In the overly replete world of Funes there were nothing but details.” Unlike Danny, Funes had lost the capacity to perceive the platonic ideal of Rocky.

Quiroga prefers the term “concept cells” to Fried’s “thinking cells,” because the cells respond not to a particular image but to the concept behind it. It’s fun to imagine the kinds of concepts a concept cell might represent. Infinity? Zero? The universe? Democracy? Entropy? War? Solipsism? God?

Does my personal paradigm, or self, depends on cells dedicated to Buddha? Bluefish? Hofstadater? Free will? Mac? Skye? Vicki? Death?

I’ll conclude with a final note, and question: In Why Machines Learn, Ananthaswamy dwells on back propagation, a crucial component of modern machine learning. Back propagation is a method whereby neural nets use past guesses to generate more and more accurate models of the world.

My current writing projects often make me reassess my past thoughts on things like the mind-body problem, quantum mechanics, the end of science, war. My compulsive re-evaluation of my convictions, I suppose, is akin to back propagation.

Ideally, back propagation should make my world-model more and more accurate. Why, then, do I feel as though I am moving further away from truth, whatever that is? Why do my ongoing attempts to understand existence leave me feeling more baffled? Even ignorant?

Final question: If an artificial intelligence doesn’t engage in this sort of self-questioning, is it truly intelligent?

Further Reading:

See my 2005 article: “Can a Single Cell Recognize Your Face?”

See this recent report on concept cells in Quanta.

See my profiles of Douglas Hofstadter and Christof Koch in Mind-Body Problems.

And see—oh, hell, just look at my list of past “Cross-Check” columns.