Created: February 23, 2020
Modified: May 04, 2020
Modified: May 04, 2020
conceptual scaffolding
This page is from my personal notes, and has not been specifically reviewed for public consumption. It might be incomplete, wrong, outdated, or stupid. Caveat lector.- Related:
- A testable prediction from this concept is that active learning---learning driven by curiosity---should be much more effective than other types, because it's driven by your conceptual scaffolding. For two propositions H1 and H2, if H1 is representable in (a currently activated part of) your conceptual scaffolding, then learning the truth of H1 you'll be able to propagate evidence up to more abstract (compressed) principles that would have predicted it. You might be able to refine concepts. You'll be able to connect this new fact with other related facts: you'll have a mental 'box' ready to accept the new knowledge. By contrast, suppose H2 is more or less totally orthogonal to your current conceptual scaffolding. Then projecting it onto your scaffolding will yield 'zero': it will literally mean nothing to you.
- Of course active learning is useful even in trivial cases because it focuses information on areas of uncertainty. If we need to learn the location of a decision boundary in [0, 1], then binary search is exponentially more effective than just seeing a bunch of labeled random inputs. (though Freund et al (1993) make the point that constructed queries may sometimes be less informative than passive observations because we lose out on information about the input distribution---for PAC-learning purposes you're better off seeing the input distribution and then querying labels for points near the boundary, essentially because you only care about getting the boundary precise when it's close to the input distribution). Is this any different? I think 'conceptual scaffolding' is maybe related to a very high-dimensional notion of decision boundaries? I should flesh this out more.
- A life conclusion
- It is much more powerful to go on coherent learning projects than to just absorb random facts. Random facts slide right off of you, because you haven't built out the scaffolding in the particular area of the fact (under a high-dimensional view of fact-space, you will almost never have built out quite the right scaffolding). Whereas a learning project builds scaffolding as you go. The facts hang on the scaffolding and, eventually, get compressed into concepts, which themselves build new scaffolding.
- I noticed this while clicking a random Pocket link. Time spent browsing the web randomly does not build scaffolding. Twitter maybe does, sometimes, if there's a concentration of conversation all around an interesting topic for a while (this happens more for me in ML twitter than politics twitter), but it's limited to very shallow scaffolding.
- Reading textbooks does build scaffolding.