Chapter 12 · Found Only by the Surface

Up to here we have followed working memory binding input with what it knows to build representations, and inscribing them into long-term memory. But being inscribed does not mean it is all used. Only when you can find and pull it out at the place you actually need it is it used at last. And that very pulling-out scarcely goes as you wish.

What we most desire in learning is to pull out what we mastered in one place at a wholly different place where the principle is similar—to apply a principle mastered in math class to a problem seen for the first time, an insight grasped in one case to an unfamiliar case. Yet one cold fact cognitive science reports most consistently is that this carrying-over scarcely happens. Why is what we most want the hardest? The answer lies in the way retrieval makes its way to long-term memory.

The Solution Not Recalled

There is a classic experiment. People are first made to read one story. A general means to strike a fortress, but driving a large army down one road sets off the mines buried in it. So he splits the army into small parts, sends them down several roads, and has them converge at once on one point before the fortress, and takes it. Then a seemingly wholly unrelated problem is given: a tumor deep in the body must be destroyed, but a ray strong enough to burn it at once burns the healthy flesh on the path it passes through. What to do?

The solution is the same as the fortress story. Fire many weak rays from all directions and gather them at the one point of the tumor. Each ray is weak enough not to harm the healthy flesh, but only where they gather does it grow strong enough to burn the tumor. "Divide a weak force into several branches and gather it at one point." The two solutions are, to the letter, the same structure. Someone who has just read that story might be expected to recall it at once.

But they did not. Of those who read the fortress story first, scarcely one in ten recalled that solution on their own before the tumor problem. Most floundered without noticing that the two problems were the same structure. The decisive thing is what came next. Told a single line—"see whether the story you just read might help"—then most solved it. It was not that they lacked the ability to apply. The ability was there; they merely could not summon it themselves.

The Index Is Hooked by the Surface

This divide—being able to solve if only you recall, yet not recalling—reveals the nature of retrieval. To recall something is for activation to spread from a now-activated node to a composite representation that has that node as a member. So retrieval succeeds only to the extent that the cue lit at this moment and the cue the representation hangs on overlap.

Then what decides which cue a representation hangs on? It is decided when it is inscribed. Since a composite representation's member is the very cue that summons it, what you wove it with when inscribing is what you can pull it out with. So inscribing does not stop at putting something into long-term memory; it settles, together, by which cue you will later pull it out. The fate of transfer is already half-decided at the moment of inscribing, because what it was woven with decides where it will light up again.

The trouble is that what is beside you and gets woven in while learning is usually the surface of that situation—the material, the objects, the context. And the cue lit at the moment of recall is likewise the surface that actually appears at that spot.

The representation stored from reading the fortress story hangs on surface nodes like army, road, fortress. But the nodes the tumor problem lights are ray, tumor, body. With not a single surface node overlapping between the two groups, there is no path for activation starting from the tumor problem to spread to the fortress representation. What of the structure the two solutions share, "gathering at one point"? That is only the form of the relation in which a composite representation binds its members—not a node that can light up, set apart, on its own. Structure is the shape of a composite representation, not an object the situation can point a finger at and light. So even when the structure is identical, if the surface differs there is no bridge for activation to cross. Surface becomes a cue; structure does not.

This also explains why the hint "see whether it might help" was an instant remedy. That one line lit "that story just now" directly, opening from outside the path that had been blocked. Had the barrier been in the ability to apply, giving the hint should not solve it. That it solves at a single hint says the barrier was not in the ability but in the index reaching that ability.

Near Transfer and Far Transfer

From this property of the index, the limit of transfer follows. In a situation whose surface overlaps with what was learned before, the relevant representation is retrieved readily—a math problem of the same type, a case of similar material. The surface resembles, so the cues overlap and activation crosses over. Such near transfer happens relatively well. But in a situation where the surface differs and only the structure is the same, even though that very structure is needed, it cannot be reached. This is why far transfer is rare.

That expertise is bound to one field and scarcely carries over to another is the same limit of the index. The structure built by someone deep in one field is stored hanging on that field's surface. A chess master's intricate formation representation hangs on the pieces and squares of the board, so where that surface is absent—like business or medicine—there is no cue to light it. So even a deep expert in one field returns to an ordinary novice in an unfamiliar field. A deeply built structure, too, is not summoned if the surface it hangs on does not appear again.

Here one question arises. What we actually want to learn is not the surface but the structure—the principle that works whatever the material. Yet that structure cannot become a cue, and so is not summoned in the new situation where it is needed. Then is there no way to make the structure a cue? If there is, how is it made?