On Eliezer’s post “The Cluster Structure of Thingspace”
I would like to request feedback on this comment, please.
I would also like to point out the cross-reference with jsalvatier’s “What are good topics for literature review prizes?”
I would like to request feedback on this comment, please.
I would also like to point out the cross-reference with jsalvatier’s “What are good topics for literature review prizes?”
I’m not actually convinced that negative examples are really necessary for learning empirical clusters in thingspace, especially if you’re just trying to teach someone a subcategory of a big class they’re already familiar with. If someone is already familiar with the concept of “bird” and you want to inform them that there is such a thing as blue jays, it may suffice to show them just a few instances of a blue jay (assuming you don’t care whether they learn the terminology). Source: this super cool paper about one-shot learning using hierarchical Bayesian models: http://www.mit.edu/~rsalakhu/papers/MIT-CSAIL-TR-2010-052.pdf
In fact you are correct: Negative examples, as in examples outside the higher-order class, are not used in the teaching of sub-classes of a “higher-order noun”. However, in discriminating between sub-classes, examples of other sub-classes serve as negatives for the sub-class you are currently teaching.
Please see chapter 11 in TOI, “Hierarchical Class Programs”, p 123.
We do care that they learn the terminology. When I said they are not accessible through ‘simple’ verbal rules like: “Listen: a bird is a small feathered flapping winged thing,” I mean not that they are deaf or completely without language that you can expand, but that they are very young children (or older children from disadvantaged backgrounds, and I can give you a real horror story from my school demonstrating how little some of these families interact with their children, see end of comment) who would mostly not process and retain even such ‘simple’-seeming rules like that.
Learners who are not yet familiar with the generalized concept of ‘higher-order nouns’, and must be shown that the way in which the verbal structure is the same for the statements “this truck is red” and “this truck is a vehicle” and the false statement “this truck is a dog” does not mean that you could find a truck that was not a vehicle in the same way that you could find a truck that’s not red, or respond to the statement “this truck is a vehicle” with “no it’s not! It’s a truck”.
A sequence for teaching the higher-order classes would start with examples of vehicles (sub-classes you are later going to teach) and differences that are as minimally different as possible (avoiding boundaries that are unclear even in the language of knowledgeable adults).
The wording of the first, modeled examples in the sequence could be like, “This is a vehicle/this is not a vehicle”. The test example wording could be, “Tell me, vehicle or not-vehicle?”
Once firm, you move on to the sequence for teaching the first sub-class.
Model perhaps a couple more, then
etc
Then in the next sequence you introduce a very different sub-class, like “boat”
etc
Continue introducing sub-classes as each becomes firm, juxtaposing members of the previously taught sub-classes with the new addition for discrimination.
Horror story: Once someone brought home the wrong kid, X, told them to watch TV, and didn’t notice that it was the wrong kid until hours later when the school finally found the free time around their frantic search for the assumed kidnapper of X (actually just the person who had come to pick up Y), and phoned Y’s family to ask why they hadn’t picked up Y yet.