D. Rumelhart, D. Norman , Modes of Learning

Rumelhart & D. Norman (1978) proposed that there are three modes of learning: accretion, structuring and tuning. Accretion is the addition of new knowledge to existing memory. Structuring involves the formation of new conceptual structures or schema. Tuning is the adjustment of knowledge to a specific task usually through practice. Accretion is the most common form of learning; structuring occurs much less frequently and requires considerable effort; tuning is the slowest form of learning and accounts for expert performance.

Restructuring involves some form of reflection or insight (i.e., metacognition) and may correspond to a plateau in performance. On the other hand, tuning often represents automatic behavior that is not available to reflection (e.g., learning procedures).

Rumelhart & Norman (1981) extended their model to include analogical processes: a new schema is created by modeling it on an existing schema and then modifying it based upon further experiences.

This didn't really make sense to me the first time I cam across it, but now it really really does. You collect evidence into conjunctive statements, test out reduction rules and refinements to find claims/formulas/functions that distribute over/apply to many clauses (which then become more atomic claims) until you have a compressed model, and the generalize those functions to see what else they apply to. Plausible structuralist accounting of learning. How would you test it on a language model?

Note: still haven't read the paper.