It's a Jungle in There: How Competition and Cooperation in the Brain Shape the Mind (19 page)

Using the terms top-down and bottom-up processing doesn’t fully illuminate the mechanisms they rely on, or at least leaves many questions unanswered. I’ve sketched a way these mechanisms might work—namely, through cooperation and competition among relevant neural representatives. A sketch is just that, however—merely an outline. In science, it’s useful to have sketches or overarching general conceptions of the ways systems work. Ultimately, though, it’s important to work out the details. Toward that aim, the next section is concerned with a phenomenon that reflects top-down and bottom-up processing in a way that has been explained in specific mechanistic terms. As it happens, the model comports with the theme of this book and, indeed, helped inspire it.

The phenomenon I’ll tell you about now is called the
word superiority
effect.
²¹
The procedure used to demonstrate the effect works like this. In one condition, participants (university students, typically) are shown a single letter, such as
d
or
k
, which is immediately covered by a visual mask, such as #$*@. Two letters are shown afterward until the participant indicates which letter
matched the letter he or she thinks was shown before, either the
d
or the
k
. People are pretty good at this. There is, after all, just one letter to perceive initially.

In another condition, the same participant is shown not one letter but four. In some cases, the four letters form a word. As in the one-letter case, the letters are masked and then the participant is shown two letters that remain on the screen until s/he indicates which letter appeared before. The test letter is shown at the same serial position as the letter whose identity is being queried, so, as in the one-letter task, there is no uncertainty about where the test letter appeared. Participants do worse in the four-letter condition than in the one-letter condition, which is not surprising since there were, after all, four letters, not just one.

The most surprising outcome of this experiment is that participants do worse in the four-letter condition than in the one-letter condition only when the four letters form a non-word. When the four letters form a
word
, participants actually do
better
than when they’re presented with a single letter. So if the four letters form “work,” participants do better at recognizing the
k
than if the original stimulus was just the letter
k
.

How can you explain this remarkable
word superiority
effect? Somehow the fact that the four letters form a word gives the letter identification process a leg up. A model that explains this outcome looks unruly, like a tangle of vines that you’d find in a jungle (
Figure 8
). The model looks so wild, so out of control, that you might wonder whether the cognitive psychologists who proposed it were in command of their senses. I assure you they were. What they realized was that out of the seeming chaos came a way of explaining the word superiority effect.
²²

According to the model, there are units tuned to particular aspects of experience—to line segments oriented in various ways, to letters, and to groups of letters comprising words. The units excite their friends (units with which they’re consistent) and inhibit their enemies (units with which they’re inconsistent). So a unit that’s tuned to a diagonal line tends to excite an
A
unit or a
K
unit, but tends to inhibit a
G
unit or an
S
unit. A unit that’s tuned to the letter
A
tends to excite word units containing that letter (words like ABLE, TRAP, TAKE, and CART) but tends to inhibit word units that don’t contain that letter (words like TRIP and TIME). The same dynamic—consistent units exciting each other and inconsistent units inhibiting each other—is expressed throughout the network.

The network I’ve described is typically shown with three levels—a feature level, a letter level, and a word level—but there are no levels as such in the network. It’s
just a matter of convenience to show word units above letter units and letter units above feature units. The levels are implicit in the connections of the units. Crucially, and happily from the point of view of making as few assumptions as possible (always a tenet of good science), the interplay of units is no different within or between putative levels.

FIGURE 8.
The interactive activation model of reading.

How does the model explain the word superiority effect? The dynamics of the model encourage constituents of words for which there is some evidence and discourages constituents of words for which there is no evidence. The constituents in turn encourage the words of which they’re a part and discourage the words of which they’re
not
a part. The constituents—the letters and the features they contain—also excite and inhibit each other depending on the history of their partnerships. Single letters have more diffuse partnerships than do entire words. Consequently, letters have a lower likelihood of being identified when presented alone than when presented with other letters in established words.

The model as a whole is called the
interactive activation
model of word perception. It provides a quantitatively confirmed account of the word superiority effect and other phenomena in reading. Its success attests to the promise of the general approach it embodies, which is that there are patterns of connectivity among neural elements whose mutual excitation and inhibition can account for key aspects of perception, including surprising results from perceptual research.

You weren’t born being able to read, so the neural network you use to do so is shaped over the course of development. So far, I’ve said relatively little about the role of development in this book, but this is a good place to say something about it, and more will be said in later chapters.

Consider how Darwin’s ideas may hold promise for the development of perception. A useful place to start is with “young pirates,” those children who wear eye patches for prolonged periods to help them see.
²³
The rationale for this therapeutic procedure is to promote normal visual development in the unpatched eye. Covering the “good” eye can help the “bad” eye improve.

Why do doctors recommend this seemingly paradoxical intervention? In children for whom the treatment is prescribed, the poorer vision of the “bad eye” may be due to a problem with moving that eye. The eye-movement difficulty may make it difficult for the two eyes to focus on the same object simultaneously, forcing the brain to choose between the two eyes’ inputs.

People who have difficulty bringing their eyes into alignment are said to have
strabismus
. In one form of strabismus, the bad eye points inward—toward the body midline. In another form, the bad eye points outward—away from the body midline. Having one eye point in a different direction than the other causes the brain to get two images rather than one. This situation forces a choice between the images—a choice that is normally unnecessary. If the choice consistently favors one eye, the brain’s ability to process visual inputs from the other eye may gradually deteriorate to the point where whatever sight was available in that not-so-good eye gets even worse, to the point, in the worst-case scenario, that that eye goes blind.

The mechanisms behind these changes were elucidated by the two neurophysiologists, David Hubel and Torsten Wiesel, whose discovery of visual feature detectors was covered in
Chapter 3
. I’ll expand on it here, both by considering visual development and, before that, by considering the role of visual feature detectors in visual perception.
²⁴

Recall from
Chapter 3
that Hubel and Wiesel found that single cells in an area of the adult cats’ brain (area V1) are tuned to particular stimuli. As discussed earlier, some of these cells are tuned to dark vertical lines standing against a light background, other cells are tuned to light vertical lines standing against a dark background, other cells are tuned to dark lines moving along a 45-degree axis, still other cells are tuned to dark lines moving along a 135-degree axis, and so on. Because the stimuli that activated the cells were simple, Hubel and Wiesel referred to these cells as
simple
feature detectors.
Hubel and Wiesel also found cells that responded to more complex ensembles of features, and they called those cells
complex
feature detectors. Hubel and Wiesel also found cells tuned to still more complicated stimulus ensembles, which they called
hyper-complex
cells.

Hubel and Wiesel were careful not to make claims about how simple, complex, and hyper-complex cells contribute to visual perception. But it was hard for others to resist suggesting that feature detectors might activate each other in such a way that their combined activity specifies real-world objects and scenes. According to this hypothesis, you can recognize a yellow Volkswagen, say, because the features for that vehicle activate relevant detectors in your brain.

This hypothesis runs into problems, however, at least if you insist that for a yellow Volkswagen to be recognized, the brain needs a yellow-Volkswagen detector.
²⁵
The hypothesis isn’t tenable because yellow Volkswagens, like virtually all objects, appear in infinitely many ways. If you needed a different yellow-Volkswagen detector for every possible yellow-Volkswagen stimulus, you’d be in trouble. In addition, if your recognition of a yellow Volkswagen depended on the firing of a yellow-Volkswagen detector, you’d have serious difficulty if that detector happened to die. “What’s that lemon-like object approaching me at high speed here on the Autobahn?” you might ask if your one-and-only cell for recognizing yellow Volkswagens became dysfunctional. Putting all your processing eggs in one basket (in one neuron) is not a good idea. More to the point, since we want to eschew top-top design in the Darwinian scheme of things, it’s not a design that’s likely to survive for very long.

If you say that single high-level feature detectors are problematic for a theory of perception, you needn’t conclude that feature detectors are useless. Sensitivity to stimulus features is crucial for perception. One way we know this is that camouflage is feature-based. The more features a target shares with its background, the harder the target is to spot. This is true for detection of soldiers in military theaters and, just as much, for detection of letters among distracters. It’s much harder to find a capital
K
among a sea of
X
s,
A
s, and
H
s than among a sea of
C
s,
O
s, and
Q
s. When a
K
is in the midst of
X
s,
A
s, and
H
s, the time to find the
K
increases with the number of distracters, but when a
K
is in the midst of
C
s,
O
s, and
Q
s, the time to find the letter hardly depends on the number of distracters. The
K
just pops out from the display.
²⁶

When visual stimuli are shown very briefly and then are masked by other stimuli, the features may mis-combine. If you’re shown a red circle and a blue triangle, for example, and these stimuli are quickly covered by a mask (a bunch of random symbols or hash marks), you may mistakenly think you
saw a red triangle and a blue circle. Such illusory conjunctions are expected if the features of the stimuli are somehow extracted from the raw stimulus and then are reassembled. Reassembling the features in the wrong way can lead to illusory conjunctions.
²⁷

The mistaken joining of features can occur in the everyday environment. If you find yourself in a dark alley and witness a stabbing, you may mistakenly believe that the person wielding the knife was a guy wearing a red beret. If at the crime scene there was actually a guy wearing a red
baseball cap
and another guy wearing a
blue
beret, the illusory conjunction of red and beret could land the wrong guy in jail.

Visual confusions like these are explained by saying that neural representatives of perceptual features get activated by the appearance of the features to which they’re tuned. Once this activation occurs, a contest follows. Features jointly activate other neural representatives to which they belong while inhibiting other neural representatives to which they don’t belong. The higher-level representatives also activate the features feeding them and inhibit the lower-level representatives that don’t. Illusory conjunctions and the other phenomena just reviewed can arise from those dynamics.

How do these results relate back to the young pirates described at the start of this section? Think back to Hubel and Wiesel, who wanted to know how feature detectors develop. They tested the hypothesis that feature detectors get strengthened if the features are supported by environmental input or get weakened otherwise. To study the fates of feature detectors, Hubel and Wiesel thought the effects of experience might be especially pronounced in young animals. They also thought that a convenient way to study the role of experience on perception might be to limit sight to just one eye.