nucleation

LIT-1056: Amortized Inference in Probabilistic Reasoning

2 min read

Properties4
StatusActive — read in full 2026-10-15 from the authors' copy on the Gershman lab site (gershmanlab.com/pubs/GershmanGoodman14.pdf, 6 pp.). The version of record is on eScholarship (escholarship.org/uc/item/34j1h7k5), whose PDF refused direct requests from this environment. The page range 517–522 is from the authors' own later citation. Worth reading as the paper that named amortized inference for cognitive science. It defines amortized inference as caching ("memoizing") sub-inferences and reusing them on related queries. It gives the first behavioural evidence for this, from two small online experiments on one Bayesian network: with the sub-query first, answers to a complex query were predictable from a person's own sub-answer (r = 0.28, against −0.11 in the reverse order), and they were faster relative to a matched baseline. For QUESTION-095 it shows that the authors treat amortization as retrieval standing in for recomputation, which supports the question's unification for internal memory only. It does not touch external storage or distributed cognition. It supplies storage cost, decomposability and frequency as untested trade-offs, but no staleness term, and no directional bias toward earlier answers, which belongs to later work (LIT-1074). For CLAIM-303 it is consistent with the trade and with frequency being one consideration among several, and its error propagation is error preserved by reuse, not staleness.
Tagscognition, probabilistic-modeling
Cited as “paper” byNOTE-886 — Gershman & Goodman — Amortized inference in probabilistic reasoning
Cited as “source” byTHEORY-339 — People's answers to a probability query depend on their own answers to an earlier query whose result it can use: a person's sub-answer predicts their later answer only when asked first, and a bias induced in a first judgment carries into a second only when the two share part of a hypothesis space, so reuse carries the earlier error with it; that reuse saves computation rests on one small response-time study and one secondary-task result

Graph View

Gershman, S. J. and Goodman, N. D. (2014). Amortized inference in probabilistic reasoning. In Proceedings of the 36th Annual Conference of the Cognitive Science Society, 517–522. https://escholarship.org/uc/item/34j1h7k5

Key takeaways

Read in full (NOTE-886). Gershman and Goodman define amortized inference as the reuse of past inferences, cached as sub-computations, to answer related later queries. They report two small online experiments in which order alone made reuse available. When a sub-query came before a complex query on a five-node Bayesian network, a person’s answer to the complex query could be predicted from their own answer to the sub-query (r = 0.28; r = −0.11 when the order was reversed). Answers to the complex query were also faster relative to a structurally matched baseline. Reuse in this paper is internal memory, and it is shown only through correlation and speed, within one task, with small effects. The paper does not measure costs. Storage cost and the choice of what to store are listed as open trade-offs. Bias, staleness and interference are not studied: reuse that may be “inaccurate” without modification, and interference and forgetting of stored inferences, appear only as questions.

Why this is here

The source of the term “amortized inference” in cognitive science, which the owner reached for: “amortized retrieval is actually a kind of amortized computation”. Read for QUESTION-095.

Backlinks