It's a Jungle in There: How Competition and Cooperation in the Brain Shape the Mind (22 page)

How did this play out? When a monkey moved its arm straight ahead toward the 12 o’clock position, a neuron that was tuned to that position fired the most, but when the same monkey moved its arm toward 11 o’clock or toward 1 o’clock, the same neuron fired less. The neuron fired still less when the monkey moved its arm toward 10 o’clock or toward 2 o’clock. It fired still less when the monkey moved its arm toward 9 o’clock or toward 3 o’clock, and so on. Each neuron had a preferred direction and an associated tuning curve (a bell-shaped function relating activation to direction, with the peak at the preferred direction), but different neurons had different preferred directions.

Just as every neuron had a preferred direction, every cardinal direction had a neuron that preferred it. So 12 o’clock was the favorite direction of one neuron, 2 o’clock was the favorite direction of another neuron, 4 o’clock was the favorite direction of another neuron, and so on. There was no large portion of the direction range that remained unpreferred by any investigated neuron. That was the second fact, a fact reminiscent of what I wrote about at the end of
Chapter 2
, that “no stone goes unturned” when it comes to natural selection.

Third, the direction of actual movement turned out to be a weighted sum of the activations of all the neurons. In other words, when all the activations were summed up, the total vector was one whose direction depended on the preferred directions of all the neurons, weighted by their activations.
¹⁶
This outcome made it possible for the arm to move in a direction that wasn’t strongly preferred by any neuron. The graded outputs of all the neurons made it possible to go in any direction, even if no neuron was specifically tuned to it.

Population coding is an elegant method for making decisions and enacting them. Not only does it not require an executive, it is also robust to degradation. If a few neurons in the “electorate” happen to fail, the system as a whole can still function. Finally, in case it’s unclear, there’s no reason why population coding can apply only to arm movement directions. It can apply to any aspect of decision-making and control for which neural “voters” can chime in to a degree reflecting their fit to the challenge at hand.

The proven power of population coding shows that skilled performance emerges from competition and cooperation among neurons. In the case of arm movements, competition is at play because neurons unsuited for needed directions get less weight than neurons that are well suited for those directions. Cooperation is at play because the collective output of all the neurons reflects all the neurons’ contributions to the ensemble.
¹⁷

Competition and cooperation among neurons are also seen in reflexes. You’re familiar with reflexes from your everyday experience. When you withdraw your hand from a hot stove, you make this motion without deliberation. You start and finish the motion in a fraction of a second.

Reflexes are vital for survival. Their neural underpinnings have been built up over eons of evolutionary time. In what sense are they reflective of an inner jungle?

Consider another kind of reflex that has been studied by neuroscientists—the response made by a cat when it feels pressure on its paw during walking. If the pressure is applied to the sole of the cat’s paw at the moment the cat steps
down
, the cat presses down a bit harder than usual. But if the pressure is applied to the sole of the cat’s paw when the cat steps
up
, the cat
raises
its foot a bit higher than usual.
¹⁸
These opposite responses, made to the same mechanical stimulus at different phases of the step cycle, occur with such short latencies that they qualify as reflexes. The differing responses reflect internal switching. The neurons activated by extra pressure on the foot cooperate with the neurons for foot
lowering
in one phase of the step cycle. But in the other phase of the step cycle, the neurons activated by extra pressure on the foot cooperate with the neurons for foot
raising
.

Cats aren’t the only creatures that display such neural nimbleness. People do too. Consider what your lips do when you make a “b” sound. Your top and bottom lips come together to interrupt the flow of air escaping from your mouth. Curious about how this works in detail, some researchers carried out an experiment whose logic was similar to the one just described. They used a specially designed torque motor to tug down gently on the lower lip of an experimental participant who agreed to say “ba, ba, ba,” over and over again. The speaker’s lower lip was pulled down painlessly and by just a few millimeters. The time of the perturbation was unpredictable for the speaker.
¹⁹

The researchers who performed the experiment wondered how the upper lip would respond when the lower lip was mechanically lowered. They found that the upper lip did just what it had to do to achieve bilabial closure. As soon as the lower lip was pulled down, the upper lip raced down after it, descending more quickly than usual in a way that had a good effect. The two lips came together just when they should have to achieve the “b.” The latency of the upper lip’s response was so short—less than 100 ms—that the response qualified as a reflex.

In another condition, the researchers didn’t have the participants say “ba ba ba.” Instead they had the participants say “fa, fa, fa.” In this case, the downward tug on the lower lip didn’t result in rapid descent of the upper lip. Instead, when the perturbation was applied, the upper lip was unaffected.

So just as the cat responded differently to the same mechanical stimulus depending on what task it was performing at the time—either lifting or lowering its foot—the human speaker in the lip-pull experiment responded differently to the same tugging on the lower lip depending on what sound s/he was trying to make—saying either “ba, ba, ba” or “fa, fa, fa.” Thus, the reflex was tuned to the current task demands in both situations.
²⁰

Such tuning of reflexes is essential for adaptive behavior. Suppose you’re in a state of high arousal when you walk down a dark alley. The slightest touch or sound can make you startle. If you’re not on guard—if, say, you’re lying on a hammock dozing beneath a pair of pines—the same touch or sound will have little effect.

A related phenomenon concerns the excitability of the muscles in the lower back and legs. The excitability of these muscles depends on what movement you’re planning to make. If you’re about to reach for an object, the excitability of your spinal reflexes related to postural control increases dramatically. Which muscles get activated before the reach depends on the properties of the reach to come.
²¹

It’s possible to model results like these with a simple neural network model that relies on cooperation and competition.
²²
Focusing exclusively on the “ba” or “fa” experiment, imagine that one neural unit propels the lower lip upward and another neural unit propels the upper lip downward. The two units inhibit each other when the sound to be made is “b” but not when the sound to be made is “f.”

Such context sensitivity can be modeled by supposing that there is another unit that either excites or inhibits an interneuron mediating the inhibition between the lower lip and upper lip. When the sound to be made is “b,” the mediating interneuron is excited, so it lets the lower lip inhibit the upper lip to an increasing degree the farther the lower lip progresses in its upward journey. However, when the sound to be made is “f,” the mediating interneuron is inhibited, so the lower lip’s activity has no effect on the upper lip. Through this balance of cooperation and competition among neurons, you can account for context sensitivity in the two-lip perturbation study. Applying the same basic logic to other behaviors makes it possible to account for context sensitivity in response to perturbations for them as well.

The fact that a simple neural-network model can account for task-dependent reflex modulation suggests that the neural machinery underlying that modulation is well established, both phylogenetically (in terms of evolution) and ontogenetically (in terms of individual development). The bottom line is that shifting alliances among neurons can account for the modifiability of reflexes.

So far in this discussion, I’ve suggested how cooperation and competition can shape motor output, but I’ve said little about learning. If the control of physical action relies on a Darwinian process, loosely conceived, the control of physical action should adapt to the demands it faces via trial-and-error learning. The reason is that natural selection is a trial-and-error process. Research on the learning of action skills bears out the expectation that trial-and-error learning plays a key role in the acquisition of action skills.
²³

Consider the development of locomotion in toddlers. These youngsters excel at exploring ways of moving. A charming example concerns toddlers in Jolly Jumpers.
²⁴
A Jolly Jumper is a baby-friendly seat suspended from a beam via a pair of elastic bands. Thanks to these big rubber bands, a baby in a Jolly Jumper can bounce up and down, provided his or her feet touch the ground. At first, the baby has no idea s/he can bounce this way. Over time, however, s/he discovers that a gentle step here, a more powerful step there, leads to jolly jumping.

Why babies find jumping so enjoyable is an interesting question to which there is no known answer, as far as I know. Setting that aside, the feature of the Jolly Jumper that makes it important here is that no instruction manual tells the baby what to do. No Mommy or Daddy stands there explaining to the infant how to get the most bang for the buck. The baby simply discovers, through trial-and-error learning, how and when to push on the floor to bounce most bountifully.
²⁵

Improving on a Jolly Jumper takes time, and during that time, babies mature. It is conceivable that maturation alone accounts for the greater efficiency of jolly jumping over time. If that were the case, however, you wouldn’t expect babies to adapt to this gizmo as quickly as they do; large gains can be observed in a matter of hours. Similarly, babies of different ages brought to the Jolly Jumper for the first time take time to get the hang of it. If jumping in this device were all a matter of maturation, older, more mature babies wouldn’t need a warm-up period. Finally, if developing the ability to jolly
jump reflected only maturation, you wouldn’t expect babies to vary as much as they do in another behavior, which, like jumping, is related to walking. That behavior is crawling.

Crawling takes many forms, as documented by Karen Adolph of New York University.
²⁶
Every infant seems to have his or her own way to crawl. Some babies crawl on their hands and feet, keeping their bellies above the ground. Other babies sit up as they scoot along, letting their heels serve as yanking hooks and their arms serve as balancing poles.

Such variability would not be expected if the development of walking were a matter of lockstep neural maturation, with all infants following the same hard-wired growth path.
²⁷
Instead, each child follows his or her own opportunistic journey. Much as species take different forms depending on the environments they occupy, the behavioral tendencies of different individuals vary, whether the behavior is crawling or some other form of action.
²⁸

Karen Adolph, the investigator who has played up the variability of crawling, has not explicitly emphasized trial-and-error learning in her description of motor development, but trial-and-error learning is compatible with her notion that motor development is a dynamic process not guided by a predetermined maturational schedule. Trial-and-error learning provides an effective way of gearing organisms to their environments, both at the level of speciation (the subject that interested Darwin) and at the level of behavior (the subject of interest here).

Given this double-duty usefulness of learning by trying and then succeeding or failing, it’s unsurprising that trial-and-error learning has proven to be such a powerful method of learning in a wide range of contexts.
²⁹
Trial-and-error learning underlies virtually all forms of perceptual-motor learning. To name just a few domains where it has proven useful, it plays a role in learning how to trace patterns viewed in a mirror, how to compensate for distortions in visual feedback caused by prisms or other optical displacements, how to make foul shots on the basketball court, how to develop track-and-field skills, and how to enhance robot learning.
³⁰

Because trial-and-error learning is prototypically Darwinian and because I want you to judge my Darwinian thesis critically, I invite you to consider the ostensive limitations of the trial-and-error approach. There are aspects of skill
learning that trial-and-error learning seems unable to explain, at least at first blush. If those difficulties are real, you might feel that the inner jungle hypothesis is doomed, and you might then want to hunt for a better theory.

One seeming limitation of trial-and-error learning is that the simple cycle implied by that phrase—try, then err, try, then err—isn’t entirely apt. That phrase makes it sound like trial-and-error learning is as rhythmically regular as a polka. It’s not. The learning of perceptual-motor skills works best when practice occurs in clumps rather than continuously. If you want to learn to touch-type, for example, you might be tempted to spend large amounts of time at the keyboard, concentrating for hours on end to learn to hit the right keys without looking at them. It turns out, however, that you can develop your typing skill more efficiently if you engage in spaced practice rather than massed practice. Spaced practice means practicing with many breaks. Massed practice means practicing for extended periods with few breaks.

Spaced practice, in general, allows for more rapid learning than does massed practice. For example, a study of learning to touch-type showed that typing students could learn more efficiently if they practiced two times a day for one hour than if they practiced one time a day for two hours.
³¹
That’s a happy outcome, isn’t it? It’s nice to discover that you can be relieved of hours of needless drudgery if you want to learn a skill. You can get a kind of free lunch by taking a break—and maybe literally going out for lunch—between practice sessions.