Dasgupta, I., Schulz, E., Goodman, N. D. and Gershman, S. J. (2018). Remembrance of inferences past: Amortization in human hypothesis generation. Cognition 178, 67–81. DOI 10.1016/j.cognition.2018.04.017
Key takeaways
Read in full (NOTE-875). In three online experiments (N = 84, 80 and 100, before exclusions), a sub- or superadditivity bias induced in a first probability judgment carried over into a second, always-packed judgment about the same scene distribution, and errors on the two queries were correlated trial by trial (r ≈ 0.34–0.67). The carry-over appeared only when the second query shared part of its hypothesis space with the first, which rules out simple numerical anchoring; it was larger when cognitive load fell on the first query, which the authors’ simulations read as reuse of a summary statistic rather than of the stored samples; and with a new cue it appeared (for subadditivity only, and weakly) when the cue’s posterior was similar to the first. The paper defines amortization as reuse of past computation bought at a cost in bias, and itself places it in the time-for-space trade. It measures no response times, so the speed benefit is assumed, not shown; efficiency is shown only by better secondary-task accuracy on reusable trials; and it gives no account of when reuse pays.
Why this is here
The owner, after the cockpit reading (LIT-1066): “amortized retrieval is actually a kind of amortized computation. there’s probably a clearer relationship here that I imagine has likely been clarified/formalized at some point in the 30 years since the cockpit essay was published.” Registered with six others for QUESTION-095.
Amortization made empirical in human inference: stored results of computation reused at a cost in bias. The closest formal reading of the owner’s point.