nucleation

LIT-1077: Memory as a Computational Resource

2 min read

Properties4
StatusActive — read in full 2026-10-15 from the publisher's typeset PDF of the version of record (TICS 25(3), 240–251), posted on the Gershman lab's publication page (gershmanlab.com/pubs/DasguptaGershman21.pdf); Unpaywall lists only the publisher's bronze copy. Worth reading as the short computer-science-to-cognition statement of memoization as a cognitive mechanism: Michie's memo-table, the space–time trade-off, top-down (whole outputs, Logan) versus bottom-up (subproblem solutions, Kalman filter, Bellman backups) reuse, and amortized inference and deep Q-networks as "distributed" memo-tables that generalize to new inputs, illustrated in arithmetic, mental rotation, inference and planning. For QUESTION-095 it shows the first strategy with the store in the head and no cost model: its equations are illustrative, when caching pays is listed as an outstanding question, retrieval cost is mentioned once (production systems' "onerous demands on storage and retrieval") and search not at all, and nothing concerns external memory, offloading or tools. It does supply both failure terms qualitatively, staleness in the strict sense (cached values not yet updated after devaluation, with offline updating as the remedy) and bias from overlap between queries, and treats query frequency as what a capacity-limited learned cache spends its capacity on. "Amortization" here means amortized inference, not amortized analysis, and the paper does not say that retrieval is computation.
Tagscognition, probabilistic-modeling
Cited as “paper” byNOTE-865 — Dasgupta & Gershman — Memory as a computational resource
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

Dasgupta, I. and Gershman, S. J. (2021). Memory as a computational resource. Trends in Cognitive Sciences 25(3), 240–251. DOI 10.1016/j.tics.2020.12.008

Key takeaways

Read in full (NOTE-865). A short review arguing that memory is used to avoid redundant computation: results are stored and reused (memoization), either whole (“top-down”) or as solutions to subproblems (“bottom-up”), and amortized inference and deep Q-networks are “distributed” memo-tables that generalize to new inputs. It illustrates this in mental arithmetic, mental imagery, probabilistic inference and planning, mostly with response times that flatten with practice and with biases that a cache explains (reuse of earlier inferences, choices based on stale cached values, reuse of policy fragments). It has equations for its examples but no cost model: the space–time trade-off is named, and when caching pays is listed as an outstanding question. Memory is internal throughout; retrieval cost is mentioned once, and offloading or external stores not at all.

Why this is here

Turned up by the citation search on QUESTION-095 (amortized retrieval as amortized computation) and registered at the owner’s request. It states that retrieval is computation, for internal memory, and by the search cites no offloading work.

The reading (NOTE-865) corrects this. The paper does not say retrieval is computation; it says memory replaces computation and treats retrieval as nearly free. What it adds is staleness in the strict sense: cached values “not yet updated” after the world changes.

Backlinks