Apophenia, take 2

After some additional consideration, I think my Apophenia post was unfair. Humans, and indeed mammals in general, are ridiculously good at detecting and utilizing correlations. But only particular types of correlations which they have been evolutionarily prepared to expect and process. Consider that darling of experimental neuroscience, the rat. Rats will learn to associate a tone with a foot-shock in only a few trials, if the tone is brief, co-terminates with the foot-shock, and if the tone predicts the shock with high reliability. If the tone is more than a few seconds long before the shock occurs, if there is a substantial gap between the tone and the shock, or if the tone occurs often without a shock, the rat will not learn the relationship. Similarly, rats can learn that water with a distinctive taste is correlated with nausea after only a single pairing, even if the nausea occurs hours after the water is consumed, but they cannot learn that lights and sounds are correlated with nausea.

You might now quite reasonably be thinking that learning to associate tones and shocks is not very impressive, and that I promised quality correlation detection. The tone-shock combinations are unnatural and thus akin to the correlations in the Apophenia post. Rats rarely encounter electrified grids hooked up to speakers in the wild. The ability to associate foods with illness is no mean feat, given the time spans involved (although it can go awry - if you've ever eaten a distinctive-tasting food while sick and later vomited, you likely found that you had lost your taste for that food. I didn't care much for lobster for most of my youth after my father brought home a special treat one evening when I had an ear infection. Cancer patients undergoing chemotherapy are often left staring down the business end of this phenomenon). But the brain (and the cortex in particular) really comes into its own when processing complex instantaneous sensory stimuli.

Consider the same rat which couldn't remember that it was going to receive a shock after hearing a bell because we nefariously inserted a five-second delay between the two stimuli. If you stick that rat in a big pool of opaque water (they use some sort of latex beads, I think) with a small platform hidden just below the surface of the water, the rat will swim around at random until stumbling upon the platform, at which point it will immediately climb up and out of the water (this is called a Morris water maze. It's generally used to test spatial memory. Note that while they're pretty good swimmers, rat's don't bathe recreationally). If you now pick the rat up, blindfold it, swing it around your head a few times to disorient it, and put it back in the pool at some random location, it will immediately swim back to the platform.

Amazing! After a single trial, the rat learned to associate the complex visual stimulus perceived at a particular location in the pool with the position of the platform. Even though this visual stimulus changes completely depending upon the direction in which the rat is facing (they don't have to approach the platform from the same direction each time). Even though the visual stimulus associated with other positions in the pool is virtually identical to that at the platform. Even though rats are not very visual animals (they rely primarily upon olfaction and tactile sensation (remember the whiskers!)). How the hell was the rat able to figure out that some subtle variation in the pattern of light falling on its retina indicates a nice dry spot to chill out, whereas almost identical patterns would leave it treading water until it drowned from exhaustion? (Of course, as the benevolent experimenter, you would rescue our friend the rat before it met its untimely demise in a kiddy-pool of milky water. Right? Right!?!?!)

Neither neuroscientists nor computer scientists have a convincing answer to this question. If you've ever tried to use voice-recognition software on a computer, you are familiar with how bad computers are at processing sensory input. The reason you've never even seen a computer vision recognition system is that they're even worse. State-of-the-art algorithms can recognize perhaps dozens of different categories of objects, but they are much less nuanced in their discriminations than a rat. For instance, most such systems are baffled by objection rotations and partial occlusion. The brain, in contrast, detects the necessary high-order correlations with such ease that you don't even realize how difficult the task is.