A Neural Network model of Cultural Evolution

Kingsley J. A. Cox; Paul R. Adams

REVIEW 3 major objections 7 minor 10 references

A Neural Network model of Cultural Evolution

T0 review · 3 major / 7 minor · reviewed 2026-07-31 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Selective social learning lets a population of simple neural agents ratchet from an easy solution to a harder but more useful one about hidden causes in their data.

desk verdict Clean ICA sims show a Light-gated flip from easy to hard fixed point, but Light is an oracle on the target IC, so the ratchet is not yet a cultural-evolution mechanism. read the letter →

arxiv 2607.24886 v1 pith:RCAAVSVQ submitted 2026-07-27 q-bio.NC

Kingsley J. A. Cox , Paul R. Adams This is my paper · ORCID

classification q-bio.NC

keywords culturalevolutionindependentcomponentanalysisneuralnetworkssociallearningHebbianratcheteffectsynapticplasticitycollectiveintelligence

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper builds a minimal neural model of cultural evolution in which single-neuron agents learn to unmix linearly combined hidden sources by nonlinear Hebbian rules. Alone, most agents settle on the easier high-kurtosis source because its basin of attraction is larger. When agents also teach one another with a selective boost called Light that strengthens learning from teachers nearer a preferred solution, the whole population can switch to the harder low-kurtosis source, even when nearly everyone starts in the easy basin. In the extreme case one pioneer already at the hard solution pulls ninety-nine others across. The result supplies an explicit synaptic account of how individual discovery plus selective social learning can produce cumulative, ratchet-like cultural improvement.

What carries the argument

Light: the factor a(cos φ)^b that scales the supervised (delta-rule) learning rate according to how close the teacher’s weight vector already lies to the preferred independent component. It lets social learning overcome the basin-size bias of unsupervised ICA dynamics.

What would settle it

Run the same one-unit ICA populations with the reported Light strengths and kurtosis ratio: if a single logistic pioneer consistently fails to convert large Laplacian majorities, or if the population never switches basins across the parameter ranges shown in the figures, the ratchet claim is false.

Watch

Extended reading notes

Core claim

With the selective factor Light that multiplies supervised learning rate by a power of the cosine of a teacher’s angle to a target independent component, a population of communicating one-unit ICA agents converges on the harder, lower-kurtosis logistic source even when the large majority begin in the larger Laplacian basin. A single agent already at the logistic solution can convert ninety-nine others, producing ratchet-like replacement of the easier solution by the designated more useful one.

Load-bearing premise

That an ad-hoc boost based on how close a teacher already is to a pre-chosen solution can stand in for real prestige, trust or utility, and that the harder solution can simply be labelled more useful.

Editorial extensions

If this is right

Selective boosting of social learning, not high-fidelity copying alone, is the key condition for cumulative cultural evolution.
Inheritance is supervised synaptic learning and variation is unsupervised Hebbian learning, both made fully explicit at the weight level.
A single pioneer at a superior solution can convert an entire population via Light-boosted teaching.
Basins of attraction of the learning dynamics can function as discrete cultural traits or memes.
The same selective-communication principle may let populations reach weight configurations inaccessible to isolated individuals once dynamics become richer.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Any reliable public cue of solution quality, not merely angle to a known component, could drive similar population ratchets in more realistic networks.
Extending the setup to nonlinear ICA or multi-layer nets would test whether selective communication still lets groups discover concepts that isolated agents essentially never find.
Prestige bias in cultural-evolution theory may be implementable as a simple activity-dependent synaptic gain control.
Once weight dynamics live on higher-dimensional spheres with saddles, intermediate ambiguous states could serve as cultural stepping-stones rather than pure discrete memes.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The load-bearing result is narrow and real: with their selective teaching factor, a population of one-unit ICA agents can switch from the larger-basin (high-kurtosis LAP) solution to the smaller-basin (LOG) one, including the extreme case of one LOG pioneer converting 99 LAP agents. Relative to their 2021/2023 papers this is the new piece—two correct fixed points of unequal difficulty instead of correct-vs-incorrect, plus the one-to-many takeover runs.

What they do well is transparency. The setup is standard one-unit ICA with cubic nonlinearity; the Appendix derivation of the separatrix from the kurtosis ratio is clean and matches the ~5:3 basin split they measure. The figures are easy to read. They are explicit that the model is toy and that unsupervised learning supplies the variation while supervised learning supplies the inheritance. That framing is useful for people who want synaptic machinery under cultural-evolution talk instead of discrete memes.

The soft spot that matters is Light. Methods define it as a(cos φ)^b where φ is the teacher’s angle to the ground-truth IC. So every learner is handed an exact, noiseless score of how close each teacher is to the pre-chosen target. With that oracle the population flow toward LOG is close to guaranteed once b is steep enough; the N=100 run needing Light(2,5.5) is consistent with overcoming 99:1 sampling odds, not with an emergent utility estimate. The Discussion’s “prestige/trust/science” gloss and the “rather arbitrarily” useful LOG label do not fix this: usefulness is assigned externally and selection is performed with access to the answer. That undercuts the claim that the model shows how a cultural ratchet can select better solutions without already knowing them. “Progressively better” and multi-step accumulation are also overstated; only a single flip between two hand-labeled attractors is shown.

Citation pattern is mostly their own prior work plus standard ICA and CCE sources; nothing misleading. Math and sims look solid for what they actually compute.

This is for people already inside computational cultural evolution or theoretical neuroscience who want a concrete synaptic sketch. It deserves a serious referee who will force the oracle issue and the scope claims into the open. I would not desk-reject it.

Referee Report

3 major / 7 minor

Summary. The manuscript presents a minimal neural-network model of cumulative cultural evolution. Single two-input, one-output agents learn by a one-unit negentropy-maximizing ICA rule (unsupervised, "U") and by a Delta-rule imitation of a randomly chosen peer (supervised, "S"), alternating with 50% probability. Two non-Gaussian sources (Laplace, excess kurtosis 3; Logistic, excess kurtosis 1.2) are mixed by an orthogonal 2x2 rotation, so two stable fixed points exist on the unit circle. The authors show analytically (Appendix) that the separatrix satisfies tan²θ* = κ1/κ2, giving a ~5:3 basin ratio favoring the higher-kurtosis (LAP) source, confirmed by 100 non-interacting agents (70/30 split). They then show that a "Light" factor, Light = a(cos φ)^b multiplying the supervised learning rate, where φ is the teacher's angle to the designated target IC, allows agents at the harder, lower-kurtosis LOG solution to convert the population — including the case of one LOG pioneer converting 99 LAP agents (Fig. 9, Light(2,5.5)). This is interpreted as a ratchet-like replacement of an easier solution by a more useful one.

Significance. If the central result holds, the paper offers a genuinely transparent instantiation of mechanisms that cultural-evolution theory usually treats as black boxes: "variation" as unsupervised Hebbian ICA learning and "inheritance" as supervised social learning, with explicit synaptic-style rules rather than meme abstractions. The Appendix derivation of the basin structure (tan²θ* = κ1/κ2) is standard but clean, and the predicted 5:3 basin ratio is quantitatively confirmed in simulation — a parameter-free check that grounds the setup. The one-pioneer-to-99-agents demonstration (Fig. 9) is a striking, falsifiable simulation result, and the authors are commendably candid about limitations (2 fixed points only, "rather arbitrarily" assigned usefulness). However, the significance for cultural evolution is currently capped by the fact that the selection signal driving the ratchet is computed from oracle knowledge of the ground-truth solution rather than from any agent-estimable quantity.

major comments (3)

[Methods §4 (Light), in relation to Figs. 5, 7, 9 and the Discussion] The load-bearing issue: φ in Light = a(cos φ)^b is 'the current angle of the teacher's weight vector with the IC' — i.e., the angle to the ground-truth independent component, which by construction is unknown to every agent (discovering it is the learning problem). Every learner therefore gates imitation by an exact, noiseless measure of teacher proximity to the pre-designated answer. With the Delta rule the effective step toward a teacher is k·a(cos φ)^b·x(y − y_T), so for sufficiently large b the population flow toward the target IC is close to analytically guaranteed; Figs. 5, 7 and 9 then confirm arithmetic rather than demonstrate a mechanism. This matters because the Discussion frames Light as 'prestige', 'trust', or an 'estimate of the value, or utility' and concludes that the results hint at 'progressive cultural shifts to increasingly superior solutions'. Those interpretations req
[Results, Figs. 2–9] Every reported result is a single exemplar trajectory. There are no repetition counts, success fractions, or variability estimates. The headline claim — one pioneer converting 99 agents (Fig. 9) — rests on one run with one hand-chosen parameter setting, Light(2,5.5), while the 4-agent analogue used Light(2,3) and the 30-agent case Light(2,3). The fact that b had to be raised from 3 to 5.5 when going from 4 to 100 agents suggests a scaling relation between required steepness and population size (plausibly because a LAP learner samples the unique LOG teacher only 1/N of supervised trials), which is arguably the most interesting quantitative question the model raises and is left unexamined. Please report: success probability over seeds as a function of b for each N; the critical steepness b*(N) if one exists; and sensitivity to the U/S mixing probability (fixed at 50% throughout and never v
[Abstract, Introduction, Discussion (ratchet framing)] The claim of 'progressively better' learning and a 'ratchet-like' process is supported only by a single transition between two pre-existing fixed points (LAP → LOG). A ratchet connotes iterated, directional improvement; here there are exactly two solutions, the 'better' one is designated by the experimenters, and no sequence of successive replacements is shown. Either demonstrate at least a two-step improvement (e.g., three sources of graded kurtosis with sequential replacement) or soften the language throughout to 'replacement of an easier solution by a designated, harder-to-find one'. This is not merely rhetorical: the Discussion's claim that the key condition for CCE is 'selective boosting of social learning by light-like factors' currently overstates what the two-attractor system can support.

minor comments (7)

[Appendix] Several typographical errors: (i) the expansion of E[y⁴] reads '6cos²θcos²θ' but should be 6cos²θsin²θ; (ii) the stationary-point condition 'κ₂sin²θ−κ₁cos²θ' is missing '= 0'; (iii) in the Example, 'the logistic distribution has κ_Lap = 1.2' should read κ_Log; (iv) the line 'θ* = 3/1.2 = 2.5' omits the arctan(√·) step applied on the next line; (v) the basin description 'that of s₁ corresponds to θ* ≤ 0 < π/2*' is garbled (should be s₂ and θ* ≤ θ < π/2). The derivation itself is correct, but these need cleanup.
[Methods §1 and Results] Fixed-point notation is inconsistent and dimensionally loose: the LAP fixed point is written as [1,1] and (1,1), but weights are normalized to the unit circle after every iteration, so the fixed points should be ±(1/√2)(1,1) and ±(1/√2)(−1,1). Also 'LP', 'LAP', and 'LO' are used interchangeably (e.g., Fig. 4 caption: 'the LO basin agent').
[Results, Fig. 3 caption] The sentence 'The plot below shows angles between the weight vectors of the agents and the LOG FP' appears twice, once with [-1,1] and once with [1,1]; the second occurrence is a copy-paste error.
[Figures generally] The figures appear as inline plots with minimal axis labeling ('time', 'cos angle') and no iteration counts, parameter values, or seed information in several captions. Please state N, Light(a,b), k, and run length in every caption, and consider plotting all agents with consistent color coding by initial basin.
[Methods §5] Matlab is mentioned but no code or data availability statement is given. Given that all claims are simulation-based, depositing the (presumably short) simulation script would make the results directly verifiable and is standard practice.
[Discussion] Typos: 'temporal depedence' (dependence); 'individual ;earning' (learning); 'most randomly-initialized agents will learn extract Laplacian signals' (missing 'to'); 'inspecific synaptic modifications' has an unbalanced parenthesis; the reference '(hyv)' appears to be a stray placeholder.
[References] Cox & Adams (2023) is cited only as an arXiv preprint; if it has since appeared in a venue, please update. Gabora (1995) lacks a venue. The Tomasello et al. (2005) citation format is inconsistent with the rest of the list.

Circularity Check

2 steps flagged · score 6.0 of 10

Light multiplies supervised updates by a(cos φ)^b with φ the teacher's true angle to a pre-designated IC, so population flow to the 'better' solution is built into the teaching rule rather than discovered from agent-estimable utility.

self definitional [Methods §4 (Light definition); cf. Results Figs. 5, 7, 9]
"Light = a(cos(φ))^b ... teachers become more effective the closer they are to the correct solution. Light can be considered as an estimate of the value, or utility, of the current solution. In this paper we will refer to the Light factors a and b as Light (a,b), so for instance Light = 2(cos(φ))^3 would be Light (2,3)."

φ is defined as the teacher's angle with the IC (the ground-truth target direction), not an agent-computable observable. The supervised update is Δw = k·Light·x(y−y_T), so the learning rate itself encodes closeness to the pre-chosen solution. Claiming that 'Light' produces selective spread of the better solution therefore restates the update rule: amplify imitation of whoever is already near the labeled IC. Convergence under strong Light (e.g. Light(2,5.5) for 1-vs-99) is near-guaranteed by construction of the bias, not an independent cultural-evolution mechanism.
self definitional [Discussion (usefulness assignment); Abstract / Intro framing of 'progressively better']
"If, rather arbitrarily, the Logistic signals are nevertheless more "useful" they can be preferentially learned using the "Light' factor. ... Here we show that with "Light" a population can learn progressively better solutions, in a ratchet-like manner."

LOG is designated 'more useful' by experimenter fiat; Light is then aimed at the LOG IC via the true-angle formula above; recovery of LOG by the population is reported as learning the more useful solution. The predicate 'better/useful' and the selection signal that enforces it are the same external choice, so the ratchet-to-better-solutions claim is definitionally satisfied once Light points at the labeled target.

full rationale

The paper's central ratchet claim is that selective social learning ('Light') lets a population replace an easier high-kurtosis (LAP) solution with a harder but 'more useful' low-kurtosis (LOG) solution, including a single LOG pioneer converting 99 LAP agents. Methods §4 defines Light = a(cos φ)^b where φ is the angle of the teacher's weight vector with 'the IC'—i.e., the ground-truth independent-component direction chosen as the target—and multiplies the Delta-rule learning rate by that factor. Teachers near the designated target therefore inject systematically larger supervised steps; with large enough b the population dynamics are a biased flow toward that known fixed point. Discussion explicitly marks usefulness of LOG as 'rather arbitrarily' assigned, then recovers LOG by aiming Light at it. The unsupervised basin structure (Appendix; kurtosis ratio → ~5:3 LAP:LOG) and the fact that intermittent supervision can move weights are independent content, but the headline result—'progressively better solutions in a ratchet-like manner'—reduces by construction to oracle-gated imitation of a pre-labeled target. This is self-definitional circularity on the load-bearing selection step, not mere self-citation of prior Cox & Adams work. Score 6: one central 'prediction' is forced by the teaching-rule definition; residual non-circular ICA dynamics remain.

Assumptions & free parameters 5 free parameters · 5 assumptions · 1 invented entities

The central claim rests on classical one-unit ICA dynamics, a supervised Delta-rule social channel, and an invented selective-teaching scalar (Light) whose parameters are chosen to produce the desired flip. Usefulness of the harder source is stipulated, not derived. No external behavioral or neural dataset anchors the parameters.

free parameters (5)

Light magnitude a = 2 (typical in figures)
Overall scale of selective social learning; set by hand per experiment (examples a=2).
Light steepness b = 2, 3, or 5.5 depending on figure
Controls how sharply teaching favors near-IC teachers; raised as high as 5.5 to flip 99 agents.
Unsupervised/supervised mix probability = 0.5
Fixed at 50% U vs 50% S each iteration without justification from data.
Learning rate k = 0.0001 (basin runs); others unstated
Set very low (e.g. 0.0001) for basin exploration; other runs unspecified precisely.
Mixing angle θ = 45 degrees
Fixed at 45° for equal source mixing; does not change relative basins but is a modeling choice.

assumptions (5)

domain assumption Observable inputs are an orthogonal linear mixture of two independent non-Gaussian unit-variance sources (LAP and LOG).
Methods §2; standard ICA generative model chosen for analytic tractability.
standard math One-unit negentropy learning with cubic nonlinearity and weight normalization has stable fixed points at the ICs, with basin boundary set by kurtosis ratio.
Appendix and citations to Hyvärinen & Oja 1998, Elliott 2012; used to interpret 70:30 split.
domain assumption Social learning is intermittent Delta-rule supervision from a randomly chosen peer’s scalar output.
Methods §3 and §5; minimal model of teaching/imitation.
ad hoc to paper The lower-kurtosis LOG source is the more useful cultural target despite its smaller basin.
Stated as arbitrary in Introduction/Results; required for the ‘better solution’ narrative.
ad hoc to paper Selective efficacy of teachers is captured by Light = a(cos φ)^b with φ the teacher’s angle to the target IC.
Methods §4; invented scalar equated loosely to prestige/trust/science.

invented entities (1)

Light (selective social-learning factor)
purpose: Break Rogers-type failure by up-weighting teachers near a designated good solution so the population can ratchet to harder ICs.
Introduced in prior author work and reused here; parameters free; no independent neural or ethnographic measurement validates the functional form.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Neural Network model of Cultural Evolution." pith.science (2026). https://pith.science/paper/RCAAVSVQ

@misc{pith2026260724886,
  author       = {Pith},
  title        = {Pith review of: A Neural Network model of Cultural Evolution},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RCAAVSVQ}},
  note         = {Machine review of arXiv:2607.24886}
}

read the original abstract

It has been proposed (Richerson and Boyd, 2008) that human intelligence is underpinned by a ratchet-like process called Cultural Evolution in which ideas, originated by individuals, can selectively spread by social learning and replace older, less fruitful ones. Useful ideas can thus accumulate beyond the lifetime of individuals. Although both social and individual learning are thought to be achieved by the selective activity-dependent adjustment of synaptic strengths in an artificial neural net-like manner, there have been relatively few attempts to incorporate neural networks into Cultural Evolution models. This has led to controversy and uncertainty about how cultural traits are created, transformed, transmitted and selected. We have constructed a transparent model of Cultural Evolution based on simple neural networks, and here we show that a population of communicating agents can learn progressively better descriptions of its environment. Specifically we generate input vectors by linearly combining hidden "causes", which agents can learn from in a nonlinear, Hebbian, manner, thus discovering synaptic weights that "unmix" the data to reveal the hidden causes. We previously showed that if agents can communicate their current estimates of hidden causes to other agents, in a selective manner we call "Light", the interacting population can learn unmixing weights under conditions where most noninteracting individuals cannot. Here we show that with "Light" a population can learn progressively better solutions, in a ratchet-like manner. Although our model is highly simplified, this simplicity allows insight into cultural evolution mechanisms that have hitherto been obscure, and provides a stepping stone to more complex, but still relatively transparent, analyses.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

10 extracted references · 7 canonical work pages

[10]

Understanding and sharing intentions: The origins of cultural cognition

Ratcheting up the ratchet: on the evolution of cumulative culture. Philos. Trans. R. Soc. Lond. B. Biol. Sci. 364, 2405–2415. https://doi.org/10.1098/rstb.2009.0052 Tomasello, M., Carpenter, M., Call, J., Behne, T., & Moll, H. (2005). “Understanding and sharing intentions: The origins of cultural cognition.” Behavioral and Brain Sciences, 28(5), 675–691 W...

arXiv 2009
[1988]

Anthropol

Does Biology Constrain Culture? Am. Anthropol. 90, 819-831,. https://doi.org/10.1525/aa.1988.90.4.02a00030. Tennie, C., Call, J., Tomasello, M.,

work page doi:10.1525/aa.1988.90.4.02a00030 1988
[1998]

Signal Process

Independent component analysis by general nonlinear Hebbian-like learning rules. Signal Process. 64, 301-313,. https://doi.org/10.1016/s0165-1684(97)00197-

work page doi:10.1016/s0165-1684(97)00197-
[2004]

Neuron Glia Biol

Experience-driven brain plasticity: beyond the synapse. Neuron Glia Biol. https://doi.org/10.1017/s1740925x05000219. Marr, D.,

work page doi:10.1017/s1740925x05000219
[2009]

Front Comput Neurosci 3, 11,

Hebbian crosstalk prevents nonlinear unsupervised learning. Front Comput Neurosci 3, 11,. https://doi.org/10.3389/neuro.10.011.2009. Cox, K.J.A., Adams, P.R.,

work page doi:10.3389/neuro.10.011.2009 2009
[2014]

Biol Cybern 108, 405–22,

Hebbian learning from higher-order correlations requires crosstalk minimization. Biol Cybern 108, 405–22,. https://doi.org/10.1007/s00422-014-0608-4. Cox, K.J., Adams, P.R.,

work page doi:10.1007/s00422-014-0608-4
[2018]

ArXiv abs/1805.08651

Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning. ArXiv abs/1805.08651. Lisberger, S.G., P.N.A.S.,

arXiv
[2021]

J Theor Biol 527, 110712,

A minimal model of the interaction of social and individual learning. J Theor Biol 527, 110712,. https://doi.org/10.1016/j.jtbi.2021.110712 Dayan, P., Abbott, L.F.,

arXiv 2021

Show all 10 references

[2023]

https://doi.org/10.48550/arXiv.2310.09175 Cox, K.J.A., Adams, P.R.,

Shedding light on social learning. https://doi.org/10.48550/arXiv.2310.09175 Cox, K.J.A., Adams, P.R.,

doi arXiv
[2024]

Epub 5;121(45):e2419747121

How neural systems transform synaptic plasticity into behavioral learning. Epub 5;121(45):e2419747121. https://doi.org/10.1073/pnas.2419747121. MacKay, D.J.C.,

doi

Pith tools

Reviewed July 31, 2026 · model on record in the stance chip above.

[1] [10]

Understanding and sharing intentions: The origins of cultural cognition

Ratcheting up the ratchet: on the evolution of cumulative culture. Philos. Trans. R. Soc. Lond. B. Biol. Sci. 364, 2405–2415. https://doi.org/10.1098/rstb.2009.0052 Tomasello, M., Carpenter, M., Call, J., Behne, T., & Moll, H. (2005). “Understanding and sharing intentions: The origins of cultural cognition.” Behavioral and Brain Sciences, 28(5), 675–691 W...

arXiv 2009

[2] [1988]

Anthropol

Does Biology Constrain Culture? Am. Anthropol. 90, 819-831,. https://doi.org/10.1525/aa.1988.90.4.02a00030. Tennie, C., Call, J., Tomasello, M.,

work page doi:10.1525/aa.1988.90.4.02a00030 1988

[3] [1998]

Signal Process

Independent component analysis by general nonlinear Hebbian-like learning rules. Signal Process. 64, 301-313,. https://doi.org/10.1016/s0165-1684(97)00197-

work page doi:10.1016/s0165-1684(97)00197-

[4] [2004]

Neuron Glia Biol

Experience-driven brain plasticity: beyond the synapse. Neuron Glia Biol. https://doi.org/10.1017/s1740925x05000219. Marr, D.,

work page doi:10.1017/s1740925x05000219

[5] [2009]

Front Comput Neurosci 3, 11,

Hebbian crosstalk prevents nonlinear unsupervised learning. Front Comput Neurosci 3, 11,. https://doi.org/10.3389/neuro.10.011.2009. Cox, K.J.A., Adams, P.R.,

work page doi:10.3389/neuro.10.011.2009 2009

[6] [2014]

Biol Cybern 108, 405–22,

Hebbian learning from higher-order correlations requires crosstalk minimization. Biol Cybern 108, 405–22,. https://doi.org/10.1007/s00422-014-0608-4. Cox, K.J., Adams, P.R.,

work page doi:10.1007/s00422-014-0608-4

[7] [2018]

ArXiv abs/1805.08651

Nonlinear ICA Using Auxiliary Variables and Generalized Contrastive Learning. ArXiv abs/1805.08651. Lisberger, S.G., P.N.A.S.,

arXiv

[8] [2021]

J Theor Biol 527, 110712,

A minimal model of the interaction of social and individual learning. J Theor Biol 527, 110712,. https://doi.org/10.1016/j.jtbi.2021.110712 Dayan, P., Abbott, L.F.,

arXiv 2021