# Chunking and Scale

Suppose I took you to an acre of prairie and told you about a family of field mice who lived there. You eye would wander over the contours of the ground and vegetation and notice a myriad of details: potential shelter in holes or under bushes; insects to eat; puddles of water to drink; a vast expanse of territory that the mice could never exhaust; and generally everything the mice would require to live happy, fulfilled, and productive lives. (Such as they are.)

But suppose instead that I led you to the same acre of prairie and told you that it was home to some grizzly bears. Your eye would consider the land in a completely different manner: no shelter; no food; no water; and far too small for giant bears to enjoy.

As is probably obvious to you, the difference is scale. Mice can life in a hole in the dirt and drink from a puddle, but bears require a cave and a river full of trout. When I show you a field and tell you about mice your attention is turns to the mouse-sized details of the terrain; when I tell you about bears you see bear-sized features. Your brain automatically chunks the whole of what it observes into components at the scale of whatever it is considering.

When you write a word, the chunks are letters -- though after we learn how to spell we rarely consciously think of words as composites of letters. When you write sentences, the chunks are words (and many people spend a great deal of time choosing the right ones). When you write paragraphs, the chunks are sentences. When you write essays, the chunks are paragraphs, connected together to serve a common theme. Essays or chapters may be put into books, books into libraries, libraries into bureaucracies, and bureaucracies into governments. And so forth. At each level, from author to librarian to bureaucrat to elected official, each person considers the same terrain at a different scale and acts appropriately.

And no one really has any idea how we do it. We know that our handling of scale is dependent on chunking, because humans can only hold five to nine items in our short-term memory. However, by combining and dividing information into chunks we can fit more into those slots. For example, when you hear a sentence your comprehension isn't limited by the number of letters in each word, but rather by the number of words in the sentence -- the letters are chunked together. When you hear a lecture you're unlikely to remember many specific sentences, but you'll be able to remember a handful of points made by the speaker. At the end of the semester you won't remember many of the lectures, but you will have incorporated the main ideas of the class into your long-term memory.

All made possible by chunking, and all completely inscrutable to artificial intelligence researchers. We have no idea how the brain makes such smart decisions about how to divide the world into chunks and then recombines those chunks into knowledge.

One of the side effects of chunking is the ability to generalize. Sometimes generalizations get a bad rap (e.g., racial discrimination), but the ability to generalize is an essential component of our intelligence. Our brains can take an observation, break it into chunks, and then later recombine some of those chunks with chunks taken from another observation. Sometimes the recombinations are faulty (e.g., superstitions) but most of the time they're incredibly useful (e.g., getting hit by a red car is just as bad as getting hit by a blue truck).

Developing artificial intelligence that can generalize from prior experience is the holy grail of the field, but it won't be accomplished without chunking, and we've got no idea how to make a computer do that well because no one knows how we do it.