Chapter 3 · Representations Grow by Binding

What is piled up in long-term memory, and in what shape? We called the unit the head handles a representation; pick up one such representation and look inside, and it turns out that most are not grains that can be split no further. The representation "red apple" is like this. It feels like one lump, but inside it the color "red" and the fruit "apple" are bound in a certain relation. Take away the color and it becomes just an apple; take away the fruit and it becomes just red. A representation is not a grain but a binding.

If a representation is a binding of smaller representations, and those smaller ones are again bindings, then the representations piled in long-term memory are densely connected, each serving as material for the others. One representation is formed by binding smaller representations and at the same time enters as material into a larger one. Drawing the whole—representations bound up and down and connected side to side—as a single network is this book's model. This network is not exactly the shape of things in the head, but seen this way, what piles up in long-term memory and how it is found and pulled out are explained as one structure, so we hold onto this picture to the end.

Primitives and Composite Representations

The network has two kinds of representation. One is a grain that can be split no further. A single stroke of a letter, some sound, some color—the bottommost representations that come straight from perception—belong here. These are called primitive nodes. They are the endpoints in the network that unravel no further downward.

The other is a composite representation made by binding those grains, or other already-made representations, together. "Red apple" is a composite representation binding "red" and "apple," and the single thought "Minsu ate an apple" is another composite representation binding "Minsu," "apple," and "eat" in a particular relation. When a more precise name is needed for this composite representation that binds several representations in one relation, it is called a hyperedge—meaning a binding that hooks not just two but three, four, several representations at once. The name does not matter. The point is the fact that a composite representation is a bundle weaving lower representations in a particular relation.

Here the way representations grow comes into view. A composite representation, once made, can itself become material for a still larger composite representation. Strokes bind into a letter, letters bind into a word, words bind into a phrase, and phrases bind into the meaning of a single sentence. A lower-layer representation enters as a member of an upper-layer one, and that upper layer becomes material for a still higher one, and so a hierarchy stacks up. This unraveling downward must stop somewhere, and that floor is precisely the primitive node that splits no further. To newly know something is for one new composite representation to be added atop already-existing ones—the network growing by one strand. This is the true nature of long-term memory broadening.

What Binds It Is the Path to Find It

The real benefit of drawing it as a network comes from this one thing: a member that makes up a composite representation is at the same time a cue that summons that composite representation.

Suppose a representation X is a composite representation binding representations A, B, and C. Then whether A lights up, or B, or C, that lighting spreads to X. "Red apple" hooks "red" and "apple" as members, and whether you see the color red or hear the word "apple," that representation wakes along with it. This phenomenon—when one representation lights up, activation spreading to representations that have it as a member—has long been treated in memory research under the name spreading activation, and has settled in as the standard framework for explaining retrieval.

From this one thing an important conclusion follows: which members you hook a representation to determines where the paths that later reach it are laid. Hook it to many cues and it can be reached by that many paths; hook it to a single isolated cue and it is reached only when that cue lights up. This is why putting something into memory is not merely "storage" but at the same time settles where to go to find it later. What you wove it together with when putting it in decides, in advance, which cue can summon it when pulling it out.

This path has its shadow too. If one representation is hooked as a member of several composite representations at once, then when it lights up, the activation splits finely into several branches and none comes up clearly. This phenomenon—the more there is attached to a single cue, the slower it becomes to pin down one particular thing with that cue—has been confirmed again and again by experiment. So if you hook a representation to a common, everyday cue, it is hard to single it out and pull it later, while if you hook it to a distinctive cue nearly bound to that representation alone, you reach it clearly at once. This is the reverse side of the principle that where you hook it determines whether you can reach it. This splitting and competition becomes the root of forgetting and interference, but that story unfolds after we have built another side of long-term memory.

Working Memory Handles the Same Network Too

We drew long-term memory as a network, but this network is not long-term memory's alone. In the previous chapter we said working memory and long-term memory pass between them the same unit, the representation, and what working memory does also happens, in the end, on top of this network. Working memory pulls a few representations from the network and lays them out (retrieval), binds several representations anew on top of that (formation), or holds onto what is floating (maintenance). Binding anew is itself growing the network by one strand, and when what is grown that way remains in long-term memory, it becomes learning.

So why one representation takes only one place in working memory while the amount of information it carries varies enormously is also explained by this network. If a representation raised into working memory is a shallow one near a primitive node, the information it holds is small; if it is a deep composite representation commanding countless lower representations, it loads the whole hierarchy beneath it, compressed into one place. Raising the single word "cat" and raising one deep composite representation, "the persimmon tree in the yard of a country house in late autumn," take the same one place in working memory, but the sizes of information they carry are beyond comparison. This difference—how deep a composite representation one lays on the same one place—varying from person to person is the seed of every story to be told about the narrow working memory.

That long-term memory is a network of representations, and that the members of a composite representation are the very cues that summon it—this is the structure of long-term memory. But structure alone is still only half. Even when the same representation hangs on the same network, some come up clearly right now while others, though clearly there, cannot be reached. We let pass earlier that what is in long-term memory and what you can reach right now are separate questions; now we look at that separate question head-on. It is the story of two forces that hang on a single representation.