How can we bring our knowledge to bear on a problem? Does this resemble how we accumulate knowledge in the first place? A thoughtful passage by David Gelernter in Mirror Worlds: or the Day Software Puts the Universe in a Shoebox…How It Will Happen and What It Will Mean explores these questions.
In your mind particulars turn into generalities gradually, imperceptibly—like snow at the bottom of a drift turning into ice. If you don’t know any general rules, if you’ve merely experienced something once, then that once will have to do. You may remember one example, or a collection of particular examples, or a general rule. These states blend together: When you’ve mastered the rule, you can still recall some individual experiences if you need to. Any respectable mind simulation must accommodate all three states. Any one of them might be the final state for some particular (perfectly respectable) mind. (Many people have been to Disneyland once, a fair number have been there a couple of times, and a few, no doubt, have been to Disneyland so often that the individual visits blend together into a single melted ice-cream puddle of a visit to Disneyland rule or script or principle or whatever. All three states are real.)
Plunge-and-squish adapts to whatever you have on hand. If there is a single relevant memory, plunge finds it. If there are several, squish constructs a modest generalization, one that captures the quirks of its particular elements. If there are many, squish constructs a sound, broad-based generalization. You may even wind up with a perma-squish abstraction, if this particular squish happens frequently enough and the elements blend smoothly together. It all happens automatically.
You need plunge and squish.
It’s worth pausing here to explain in a little more detail plunge and squish. Plunge is when you take a new case—”one attribute or many attributes, doesn’t matter”—and plunge it into the memory pool. “The plunged-in case attracts memories from all over: The ‘force fields’ inside the system get warped in such a way that every stored memory (every case in the database) is re-oriented with respect to the plunged-in “bait.” The most relevant memories approach closest; and the less-relevant ones recede into the distance.” Squish, on the other hand, means “to look at the closest cases that are attracted by a plunge, and compact them together into a single ‘super case.’ We take all these nearby memories (in other words) and superimpose them.”
One more point: Whatever stack of memories you have on hand, you can cut the deck in a million ways. You can reshuffle it endlessly. You can, if you need to, synthesize a general rule at a moment’s notice. You see an asphalt spreader on the next block. You develop an expectation: The next block will smell like [the smell of fresh asphalt…}. What happened—did you wrack your brain for that important general principle, squirrelled away for just such an occasion—fact number three million twenty-one thousand and seven—fresh asphalt usually smells like…? Or did you synthesize this rule by doing a plunge-and-squish on the spot?
Clearly you can cobble together an abstraction, a category or an expectation at a moment’s notice. You can create new categories to order whenever they are needed. (Unpleasant vacations? Objects that look like metal but aren’t?…) Any realistic mind simulation must know how to do this.
Gotta have plunge; gotta have squish.
And so we arrive, finally, at two radically different pictures of the mind. In the mind-map view, there is a dense intertwined superstructure of categories, rules and generalizations, with the odd specific, particular fact hanging from the branches like the occasional bird-pecked apple. In the plunge-and-squish view, there are slowly-shifting, wandering and reforming snowdrifts instead, built without superstructure out of a billion crystal flakes—a billion particular experiences. New experiences sift constantly downwards onto the snowscape and old ones settle imperceptibly into ice-clear universal, and the whole scene is alive and constantly, subtly changing.
It’s too soon to say which view is right. Both approaches need a lot more work. Both have produced interesting results. …