2012 : WHAT IS YOUR FAVORITE DEEP, ELEGANT, OR BEAUTIFUL EXPLANATION? [1]

robert_sapolsky's picture [5]
Neuroscientist, Stanford University; Author, Behave

The Simpleton Ant and the Intelligent Ants

The obvious answer should be the double helix. With the incomparably laconic "It has not escaped our notice….," it explained the very mechanism of inheritance. But the double helix doesn't do it for me. By the time I got around to high school biology, the double helix was ancient history, like pepper moths evolving or mitochondria as the power houses of the cell. Watson and Crick—as comforting but as taken for granted as Baskin and Robbins.

Then there's the work of Hubel and Wiesel, which showed that the cortex processes sensations with a hierarchy of feature extraction. In the visual cortex, for example, neurons in the initial layer each receive inputs from a single photoreceptor in the retina. Thus, when one photoreceptor is stimulated, so is "its" neuron in the primary visual cortex. Stimulate the adjacent photoreceptor, and the adjacent neuron activates. Basically, each of these neurons "knows" one thing, namely how to recognize a particular dot of light. Groups of I-know-a-dot neurons then project onto single neurons in the second cortical layer. Stimulate a particular array of adjacent neurons in that first cortical layer and a single second layer neuron activates. Thus, a second layer neuron knows one thing, which is how to recognize, say, a 45 degree angle line of light oriented. Then groups of I-know-a-line neurons send projections on to the next layer. 

Beautiful, explains everything—just keep going, cortical layer upon layer of feature extraction, dot to line to curve to collection of curves to…….until there'd be the top layer where a neuron would know one complex, specialized thing only, like how to recognize your grandmother.  And it would be the same in the auditory cortex—first layer neurons knowing particular single notes, second layer knowing pairs of notes….some neuron at the top that would recognize the sound of your grandmother singing along with Lawrence Welk.

It turned out, though, that things didn't quite work this way. There are few "Grandmother neurons" in the cortex (although a 2005 Nature paper reported someone with a Jennifer Aniston neuron). The cortex can't rely too much on grandmother neurons, because that requires a gazillion more neurons to accommodate such inefficiency and overspecialization. Moreover, a world of nothing but grandmother neurons on top precludes making multi-modal associations (e.g., where seeing a particular Monet reminds you of croissants and Debussy's music and the disastrous date you had at an Impressionism show at the Met. Instead, we've entered the world of neural networks.

Switching to a more mundane level of beauty, consider the gastrointestinal tract. In addition to teaching neuroscience, I've been asked to be a good departmental citizen and fill in some teaching holes in our core survey course. Choices: photosynthesis, renal filtration or the gastrointestinal tract. I picked the GI tract, despite knowing nothing about it, since the first two subjects terrified me. Gut physiology turns out to be beautiful and elegant amid a huge number of multi-syllabic hormones and enzymes. As the Gentle Reader knows, the GI tract is essentially a tube starting at the mouth and ending at the anus. When a glop of food distends the tube, the distended area secretes some chemical messenger that causes that part of the tube to start doing something (e.g., contracting rhythmically to pulverize food). But the messenger also causes the part of the tract just behind to stop doing its now-completed task and causes area just ahead to prepare for its job. Like shuttling ships through the Panama Canal's locks, all the way to the bathroom.

Beautiful. But even bowelophiles wouldn't argue that this is deep on a fundamental, universe-explaining level. Which brings me to my selection, which is emergence and complexity, as represented by "swarm intelligence."

Observe a single ant, and it doesn't make much sense, walking in one direction, suddenly careening in another for no obvious reason, doubling back on itself. Thoroughly unpredictable.

The same happens with two ants, a handful of ants. But a colony of ants makes fantastic sense. Specialized jobs, efficient means of exploiting new food sources, complex underground nests with temperature regulated within a few degrees. And critically, there's no blueprint or central source of command—each individual ants has algorithms for their behaviors. But this is not wisdom of the crowd, where a bunch of reasonably informed individuals outperform a single expert. The ants aren't reasonably informed about the big picture. Instead, the behavior algorithms of each ant consist of a few simple rules for interacting with the local environment and local ants. And out of this emerges a highly efficient colony.

Ant colonies excel at generating trails that connect locations in the shortest possible way, accomplished with simple rules about when to lay down a pheromone trail and what to do when encountering someone else's trail—approximations of optimal solutions to the Traveling Salesman problem. This has useful applications. In "ant-based routing," simulations using virtual ants with similar rules can generate optimal ways of connecting the nodes in a network, something of great interest to telecommunications companies. It applies to the developing brain, which must wire up vast numbers of neurons with vaster numbers of connections without constructing millions of miles of connecting axons. And migrating fetal neurons generate an efficient solution with a different version of ant-based routine.

A wonderful example is how local rules about attraction and repulsion (i.e., positive and negative charges) allow simple molecules in an organic soup to occasionally form more complex ones. Life may have originated this way without the requirement of bolts of lightening to catalyze the formation of complex molecules.

And why is self-organization so beautiful to my atheistic self? Because if complex, adaptive systems don't require a blue print, they don't require a blue print maker. If they don't require lightening bolts, they don't require Someone hurtling lightening bolts.