Copying complex descriptions. It is easy to imagine ways to record a simple list of symbols or properties, but I have never seen any plausible schemes for how a brain could quickly make copies of structures with more complex connections. That is why this section proposed to use a sequential, difference engine-like scheme.[164]
Learning by building “Knowledge-Lines”
Suppose that you’ve just had a ‘good idea’ or solved a certain hard problem P. What, then, should you learn from this experience? One thing you could do is to construct a new rule: If the problem you’re facing is like problem P, Then try the solution that worked on P. That solution will probably help you with problems that closely resemble P—but it is unlikely to help with less similar problems.
However, instead of just recording your solution to P, you could make a record of the Way To Think that used were using when you found that solution! Generally, one would expect this to help with a much wider range of other problems. But how you could make copy of anything like your entire state of mind? Clearly, that would never be practical—but you might get a good approximation to it if you could, later, re-activate a substantial portion of the resources that were active when you solved problem P.
This suggests that to remember the method you used to find the solution to P, you could simply construct a new Selector that activates that set of resources. We call this kind of structure a “K-line,”
One can see such a K-line as a sort of ‘snapshot’ of a mental state because, when you later activate it, that will put you into a similar state, which should help to solve other problems like P. Here is an analogy that illustrates how K-lines work:
Kenneth Haase: “You want to repair a bicycle. Before you start, smear your hands with red paint. Then every tool you need to use will end up with red marks on it. When you’re done, just remember that ‘red’ means ‘good for fixing bicycles.’ If you use different colors for different jobs, some tools will end up marked with several colors. […] Later, when you have some job to do, just activate the set of tools with the right color for that kind of job, and the resources that you’ve used for such jobs then become available.” [See SoM 8.1]
This way, for each kind of problem or job, your K-lines can fill your mind with ideas that might be relevant—by getting you into a mental state that resembles one that, in the past, helped with a similar job.
Student: How could this help you do anything new, if it only re-uses a way to think that you already knew how to use—when you used it to solve that old problem?
Whenever you work on a novel problem, you may start up several K-lines at once, and this is likely to activate some resources that conflict with one another. Then, this will lead to a cascade in which some resources will get suppressed and some other resources will be aroused—and now you’ll be using a somewhat new set of resources, and your state of mind won’t be quite the same as any that you have been in before. Thus, every new situation is likely to lead to a somewhat novel Way to Think—and if you make a ‘snapshot’ of that, you will now have a K-line that differs from all of your previous ones.[165]
Student: I see how this new K-line could be used as a Selector for a new Way to Think. But how would you build a new Critic to recognize when to activate it?
If we want to use that new K-line for problems like P, then a simple such Critic would recognize some combination of features of P.
However, such a Critic will rarely be useful if it requires all the features of P, because then it will not recognize situations that are slightly different that P. Also, each new K-line should include only resources that actually helped.
Student: I see what you mean. Suppose that when you were fixing that bicycle, at some point you tried to use a tool that happened to make the problem worse. It wouldn’t be good to paint that tool red because, then, later you would waste more time again.
This suggests that when we make new Selectors and Critics—or more generally, whenever we learn—we should try to make sure that what we learn will mainly consist of things that are likely to help. The following section discusses some processes that might help to ensure that what we learn will be relevant in future times.
§8-5. Credit-Assignment
When first we met Carol in chapter §2, she learned to use spoons for moving fluids. But then we asked about which aspects of her several attempts should get credit for her final success:
Should her learning include which shoes she wore, or the place in which those events occurred, or whether the weather was cloudy or clear? Which of the thoughts she was thinking then should be recorded in what she remembers? What if she smiled while using that fork, but happened to frown when using that spoon? What keeps her from learning irrelevant rules like, “To fill a cup, it helps to frown”?[166]
Some early ideas about how animals learn were based on schemes in which each reward for a success will cause a small “reinforcement” of some connections in the animal’s brain—while every disappointment will cause a corresponding weakening. In simple cases, this can lead a brain to select the right features to recognize. [See §§§Reinforcement] However, in more complex situations, such methods will not work so well to find which features are relevant—and then we’ll need to think more reflectively.
Some other theories about how learning works assumed that this consisted of making and storing new If-Do reaction-rules. This could be one reason why Jack’s dog in §84 needed so many repetitions: perhaps, each time that dog attempted that trick, it made a small change some If or some Do—but then, it only recorded that change in the case that it got a reward.
That might work well in a simple case. However, in more complex situations, this kind of learning is likely to fail if the If of any new If-Do rule turns out to describe too few details (because then that rule will be applied too recklessly)—or if that If includes too many details (because then it may never apply again, because no two situation are ever exactly the same.) The same applies to the Do of that rule—and so, each If and Do must be just abstract enough to make it apply to ‘similar’ situations—but not to too many dissimilar ones. Otherwise, Jack’s dog might need a new rule for every posture that it might be in. All this means that those old ‘reinforcement’ schemes might work well enough for some animals, but it seems unlikely they could be help much to explain how humans learn more complicated things.
165
The K-line idea was first developed in [Ref: Plain Talk] and [Ref: K-lines]. Chapter 8 of SoM describes more ideas about what might happen when K-lines conflict.
166
Perhaps she used that facial expression to help her maintain her concentration. If this became part of her subsequent skill, it could later be hard to eliminate.