Semantic and episodic memories in a predictive coding model of the neocortex

Fr\'ed\'eric Alexandre (Mnemosyne); Lucie Fontaine (Mnemosyne)

REVIEW 3 major objections 4 minor 29 references

Semantic and episodic memories in a predictive coding model of the neocortex

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A predictive coding model of the neocortex can recall individual training images when overfitted to a small minibatch, but loses this episodic recall when trained on the full dataset, supporting the Complementary Learning Systems division o

desk verdict A clear PCN memory demonstration with a nice semantic-bar artifact, but the central training-set-size comparison is confounded by the PC/iPC switch and purely qualitative recall evidence. read the letter →

arxiv 2509.01987 v1 pith:3R2QIJZ6 submitted 2025-09-02 cs.LG cs.NE

Lucie Fontaine (Mnemosyne) , Frédéric Alexandre (Mnemosyne) This is my paper · ORCID

classification cs.LGcs.NE

keywords predictivecodingepisodicmemorysemanticcomplementarylearningsystemsneocortexhippocampusauto-associativegenerativemodel

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 asks whether the neocortex, usually thought to store semantic knowledge in dense overlapping representations, can also store individual episodic memories. Using a three-layer predictive coding network trained on handwritten digits, it shows that the network can recall specific training images when trained on a single minibatch of 64 examples—but this is overfitting, and the same network trained on the full training set reconstructs images well yet fails to recall the details of individual examples. The authors read this as evidence that a semantic learning system can encode a limited number of episodes by memorizing them, but that a separate sparse, pattern-separated hippocampal system is needed once the number of episodes grows. If correct, this supports Complementary Learning Systems theory over the idea that the neocortex alone can serve both memory functions.

What carries the argument

The central object is a three-layer predictive coding network with a hierarchical Gaussian generative model, local prediction-error units, and inference and learning by gradient descent on variational free energy. A replay protocol isolates the top-level cause and zeroes the input prediction error to regenerate an image; an auto-associative recall task presents a partial input and lets the input layer converge under the same error dynamics. The comparison between a 64-example training run and a full-dataset training run is the instrument that exposes the trade-off between memorization and generalization.

What would settle it

Train the same predictive coding variant, with identical hyperparameters, on the small minibatch and on the full dataset; if the small-data model and the full-data model show the same recall fidelity, or if recall does not decline as the training set grows, the central claim fails. A continuous sweep of training-set size under a single algorithm would settle whether the loss of episodic recall is caused by dataset size or by the change to incremental predictive coding.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that episodic recall in a predictive coding neocortex model is inversely related to the breadth of semantic learning: a model overfitted to 64 examples can complete partial versions of those examples, while a model trained on about ten thousand examples converges to a general code for the digit classes and loses the specifics of individual images, even though its reconstructions improve. The mechanism is the density and overlap of the learned representations: with few examples, the weights become a collage of those specific images; with many examples, individual traces are blended into class-level statistics. The paper concludes that dense o

Load-bearing premise

The two experiments differ not only in the number of training examples but also in the learning algorithm (plain predictive coding versus incremental predictive coding), so the loss of recall could be caused by the algorithmic change rather than by dataset size alone.

Editorial extensions

If this is right

If the claim is right, a single neocortex-like learning system cannot serve as a full episodic store; some form of sparse, fast-learning hippocampal system is required for detailed recall of many individual experiences.
Models of consolidation that assume episodic traces are transferred to the neocortex need to specify how those traces survive interference; this result places a capacity limit on neocortical episodic storage.
In AI, generative models trained to convergence on large datasets should be expected to reconstruct class statistics rather than individual training examples; auto-associative recall of specific items is a small-data overfitting phenomenon.
The result operationalizes episodic memory as pattern completion of partial inputs and gives a testable measure—recall fidelity versus training-set size—for comparing neocortical models.
The replay protocol used here, inferring the top-level cause and then regenerating the lower layer with input error silenced, could serve as a benchmark for studying memory consolidation in predictive coding networks.

Reading between the lines

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

The paper does not vary the training-set size continuously under a single algorithm, so its central trade-off remains a two-point comparison; an obvious extension is a size sweep (for example 64, 256, 1k, 10k examples) using the same predictive coding variant in all conditions.
The result suggests a concrete boundary: the number of episodes a dense overlapping system can store may scale with representational width and capacity, and increasing the hidden layer could push the recall-versus-generalization crossover—an untested consequence the authors themselves flag.
If the recall failures come from pattern-blending in a low-dimensional latent space, then higher-dimensional or more structured latent codes might preserve more episodic detail while still generalizing; this is an architectural hypothesis the paper does not test.
The replay step of silencing input error and regenerating from the top-level cause is a candidate mechanism for 'semanticized' episodic recall; testing whether replayed images degrade systematically with training size would connect this model to the clinical observation that remote memories lose episodic detail.

Share X Bluesky LinkedIn Reddit HN

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

Colleague,

The paper is worth a look as a clean, readable demonstration that a small PCN trained on 64 MNIST digits can autoassociate those digits, while the same architecture trained on ~10k digits reconstructs generic 4s and 7s. That contrast, including the recurring 'semantic bar' artifact, is a genuinely nice illustration of how dense overlapping representations can store a few episodes by overfitting. The derivation of the update rules in Section III-A is standard but clearly presented, and the replay/recall protocol is sensible. For that, credit is due.

The problem is that the two experiments do not isolate training set size. Experiment 1 uses PC with beta=1e-4; Experiment 2 uses iPC with beta=1e-5, and the authors themselves say iPC is required for stability on the full dataset. That changes both the optimization trajectory and the effective update schedule, so the loss of exemplar recall could be an artifact of the algorithm switch or the lower learning rate, not the amount of data. The paper needs to run both algorithms on both dataset sizes (or at least on a medium-sized set) to make its central claim. This is not a minor caveat; the abstract and Discussion are built on that comparison.

Also, recall is assessed only by eye on ten images. No scalar metric like MSE or, better, hit rate on a larger recall set. The training curves are 'not shown' for both experiments, so we cannot evaluate the overfitting claim quantitatively. The leap from 4s and 7s to the CLS neuroscience conclusion is large; it is a suggestion, not a demonstration.

The paper is not incoherent; it is a clear empirical study with an interesting observation. But the central inference is confounded, and the evidence is qualitative. Send it to review if the venue can handle a constructive major revision; a serious referee could push the authors to fix the comparison and add metrics. I wouldn't cite it in its current form, but I'd read a revised version.

Referee Report

3 major / 4 minor

Summary. The paper studies whether a three-layer predictive coding network (PCN) can support episodic-like recall, as a challenge to Complementary Learning Systems (CLS) theory. The authors derive inference and learning rules for a PCN with dimensions [784,35,2], present a mapping of the network onto LGN/visual cortex/entorhinal cortex, and define replay and associative-memory (AM) recall procedures. In Experiment 1, a PCN is trained on a single 64-image minibatch of MNIST digits 4 and 7; the model overfits, reconstructs training images well, and appears to recall the bottom half of ten training images when given the top half. In Experiment 2, an incremental PC (iPC) variant is trained on the full 10097-image training set; the model generalizes better but, according to visual inspection, recalls individual training images less faithfully. From this contrast the paper concludes that dense, overlapping neocortical-style representations can encode a limited number of individual episodes by overfitting but that this ability is lost as the number of examples grows, thereby supporting the need for a separate hippocampal episodic system. The paper also discusses replay, t-SNE visualizations of learned representations, and related work on predictive coding and associative memory.

Significance. If the central claim were established, the paper would make a useful contribution to the CLS debate: it would provide a concrete demonstration that a neocortical-style generative model can exhibit episodic-like recall in a small-data regime and lose that ability in a large-data regime, suggesting a computational reason for sparse, pattern-separated hippocampal representations. The derivation of the update rules in Section III-A is standard and appears correct, and the implementation choices—hand-derived gradients, no autodiff, explicit replay and recall protocols—are transparent. The paper also makes a falsifiable qualitative prediction: recall fidelity should decrease with training-set size for a fixed predictive-coding architecture. However, as detailed below, the current experiments do not isolate training-set size, and the absence of quantitative recall metrics makes the central claim difficult to evaluate. The paper is therefore significant in scope but not yet supported at the level of evidence required.

major comments (3)

[Table I and Section III-C] The central comparison between Experiment 1 and Experiment 2 confounds training-set size with model variant, learning rate, and update rule. Experiment 1 uses PC with beta=1e-4 on a 64-image minibatch; Experiment 2 uses iPC with beta=1e-5 on the full 10097-image training set. The text itself states that iPC is required for stability on the whole dataset. Since iPC updates weights during inference rather than after inference converges, it can follow a different optimization trajectory and may settle in different attractors even if the fixed points coincide. Thus the observed loss of recall in Experiment 2 could be due to the iPC rule, the lower learning rate, or the interaction of these with the larger dataset, rather than to the number of training examples per se. The conclusion that training-set size causes the loss of episodic recall is not established unless the comparison is made wit
[Section IV (Figures 6 and 9)] Recall performance is assessed only by visual inspection of ten hand-picked training images. The text states that in Experiment 1 the episodic memories were 'successfully recalled' and in Experiment 2 'not recalled as well,' but no scalar metric, error bar, or statistical comparison is provided. A per-image measure (e.g., MSE/SSIM between recalled and original bottom halves, or fraction of recalled images closer to the correct original than to the class prototype) over the full training set would be needed to support the claim that recall capability is lost. Without such a measure, the central qualitative contrast is not quantitatively specified, and the hand-picked examples in Figures 6 and 9 cannot rule out that the apparent difference is due to image difficulty or selection effects.
[Section III-C and Section IV] The paper reports convergence of prediction errors but twice refers to plots that are 'not shown' (training/validation prediction-error curves for both experiments). Given that the interpretation of Experiment 2 rests on the claim that training converged to a good fit while recall degraded, these curves are load-bearing evidence. Without them, the reader cannot verify that the two models were trained to comparable convergence criteria, nor assess whether the 'overfitting' in Experiment 1 is actually reflected in a divergence between training and validation errors. The paper should include the training curves or, at minimum, report final error values for all layers and datasets.

minor comments (4)

[Abstract and throughout] Several typographical errors should be corrected: 'exemples' (Abstract), 'attactors' (Section I-B), 'questionned' (Section I-B), 'sytems' (Section I-A1), 'phenomeon' (Section I-A2), 'hierachical' (Section II), 'representions' (Section VI), 'fom' (Section II), 'independant' (Section II).
[Section III-C] The sentence 'Except for the content of the training dataset, the two experiments are performed in the same way' is directly contradicted by Table I, which shows different algorithms (PC vs iPC) and different learning rates (1e-4 vs 1e-5). Please revise to state explicitly which variables are held constant and which are not.
[Section III-B] The replay procedure sets the precision of the input layer to zero (or equivalently xi_1=0) but the text does not explain how this is implemented in the discrete update equations (4)-(7). Clarify whether the xi_1 term is simply omitted from the phi_2 update during replay.
[Figure 4 caption] The caption says the first two levels correspond to u and phi_2 and are visualized with t-SNE, but the text says the bottom plot is the original image. Please clarify what each subplot shows and how t-SNE was applied (e.g., per-layer, on which representation).

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity detected: recall behavior is an emergent training outcome, not a fitted input or self-cited premise.

full rationale

The paper's central claim is that a predictive coding network can recall individual examples when trained on a small minibatch but loses this ability when trained on the full dataset. This claim is supported by direct experiments rather than by defining the result into the model. The update rules in Section III-A are derived from the variational free energy, and the experimental outcomes in Figures 6 and 9 are observed behaviors of the trained networks, not quantities constructed to equal the fitted hyperparameters. The use of iPC in Experiment 2 is borrowed from an external reference [24] and is an algorithmic choice, not a self-citation by the authors (the reference list contains no work by Fontaine or Alexandre). No uniqueness theorem or ansatz is smuggled in from prior work by the same authors. The main weakness is experimental confound: Experiment 1 uses PC with beta=1e-4 while Experiment 2 uses iPC with beta=1e-5, so the comparison of recall performance across experiments does not isolate training set size. This is a correctness/validity concern, not a circularity concern, because the conclusion is not equivalent to its inputs by construction. The absence of a scalar recall metric further weakens the quantitative claim, but again this is an evidence-quality issue rather than circular reasoning. Overall, the derivation chain is self-contained and the finding is not imposed by definition, fitted parameters, or self-citation.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The model relies on hand-chosen architecture and hyperparameters, on an assumption that the replay protocol is biologically faithful, and on an assumption that iPC and PC are interchangeable. No new physical or computational entities are postulated; the layers are mapped to existing structures LGN, visual cortex, and entorhinal cortex.

free parameters (6)

layer dimensions [784,35,2] = 784, 35, 2
Fixed bottleneck architecture; width 35 and two latent dimensions shape capacity for memorization.
activation function f = tanh
Chosen from empirical trials; nonlinearity affects reconstruction and recall geometry.
batch size = 64
In experiment 1 this is the entire training set; memorization depends directly on this number.
inference iterations T = 50
Number of E-step updates; convergence of replay and recall depends on it.
inference rate alpha = 0.01
Hand-tuned; affects convergence of latent estimates.
learning rate beta = 1e-4 (PC), 1e-5 (iPC)
Hand-tuned; differs between the two experiments, contributing to the PC/iPC confound.

assumptions (4)

domain assumption A three-layer Gaussian hierarchical generative model with identity covariances and standard normal prior is an adequate model of neocortical semantic learning.
Section III-A defines the model; the brain mapping and all conclusions depend on this adequacy.
domain assumption Replay can be implemented by setting input precision to zero, following Barron et al.
Section III-B; if this precise manipulation is not a faithful model of hippocampal-neocortical replay, the replay observations do not speak to episodic memory.
domain assumption MNIST digits 4 and 7 are sufficient proxies for natural episodic and semantic content.
Experiments use only this subset; generalizing to cortex is an extrapolation.
ad hoc to paper Incremental PC has the same fixed points as PC, so using iPC in experiment 2 does not alter the comparison.
Section III-C switches to iPC for stability without showing that the recall difference is independent of this algorithmic change.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Semantic and episodic memories in a predictive coding model of the neocortex." pith.science (2026). https://pith.science/paper/3R2QIJZ6

@misc{pith2026250901987,
  author       = {Pith},
  title        = {Pith review of: Semantic and episodic memories in a predictive coding model of the neocortex},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3R2QIJZ6}},
  note         = {Machine review of arXiv:2509.01987}
}

read the original abstract

Complementary Learning Systems theory holds that intelligent agents need two learning systems. Semantic memory is encoded in the neocortex with dense, overlapping representations and acquires structured knowledge. Episodic memory is encoded in the hippocampus with sparse, pattern-separated representations and quickly learns the specifics of individual experiences. Recently, this duality between semantic and episodic memories has been challenged by predictive coding, a biologically plausible neural network model of the neocortex which was shown to have hippocampus-like abilities on auto-associative memory tasks. These results raise the question of the episodic capabilities of the neocortex and their relation to semantic memory. In this paper, we present such a predictive coding model of the neocortex and explore its episodic capabilities. We show that this kind of model can indeed recall the specifics of individual examples but only if it is trained on a small number of examples. The model is overfitted to these exemples and does not generalize well, suggesting that episodic memory can arise from semantic learning. Indeed, a model trained with many more examples loses its recall capabilities. This work suggests that individual examples can be encoded gradually in the neocortex using dense, overlapping representations but only in a limited number, motivating the need for sparse, pattern-separated representations as found in the hippocampus.

Figures

Figures reproduced from arXiv: 2509.01987 by the authors.

**Figure 2.** Experiment 1: Reconstructions for a mini-batch of images from the [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Experiment 1: Visualization of weights θ1 as 35 images of size 28 × 28. As a result, replayed inputs corresponding to these high-level representations might be ambiguous. This is what we see in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: Experiment 1: Visualization of the hierarchical representations of the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Experiment 1 : Replayed inputs (right) for a mini-batch of training [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 7.** Figure 7: Experiment 2 : Reconstructions for a mini-batch of images from the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 9.** Figure 9: Experiment 2: AM task. Ten images from the training set (right [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 22 canonical work pages

[1]

A generative model of memory construction and consolidation,

E. Spens and N. Burgess, “A generative model of memory construction and consolidation,” Nature Human Behaviour , pp. 1–18, 2024

work page 2024
[2]

Off-line replay maintains declarative memories in a model of hippocampal-neocortical interactions,

S. K ´ali and P. Dayan, “Off-line replay maintains declarative memories in a model of hippocampal-neocortical interactions,” Nature Neuroscience, vol. 7, no. 3, pp. 286–294, 2004

work page 2004
[3]

What learning systems do intelligent agents need? complementary learning systems theory updated,

D. Kumaran, D. Hassabis, and J. L. McClelland, “What learning systems do intelligent agents need? complementary learning systems theory updated,” Trends in cognitive sciences , vol. 20, no. 7, pp. 512–534, 2016

work page 2016
[4]

Retrograde amnesia and memory consolidation: a neurobiological perspective,

L. R. Squire and P. Alvarez, “Retrograde amnesia and memory consolidation: a neurobiological perspective,” Current Opinion in Neurobiology, vol. 5, no. 2, pp. 169–177, 1995. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0959438895800239

arXiv 1995
[5]

Anterograde and retrograde amnesia in rats with dorsal hippocampal or dorsomedial thalamic lesions,

G. Winocur, “Anterograde and retrograde amnesia in rats with dorsal hippocampal or dorsomedial thalamic lesions,” Behavioural Brain Research, vol. 38, no. 2, pp. 145–154, 1990. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0166432890900124

work page arXiv 1990
[6]

The primate hippocampal formation: Evidence for a time-limited role in memory storage,

S. M. Zola-Morgan and L. R. Squire, “The primate hippocampal formation: Evidence for a time-limited role in memory storage,” Science, vol. 250, no. 4978, pp. 288–290, 1990. [Online]. Available: https://www.science.org/doi/abs/10.1126/science.2218534

work page doi:10.1126/science.2218534 1990
[7]

Three cases of enduring memory impairment after bilateral damage limited to the hippocampal formation,

N. L. Rempel-Clower, S. M. Zola, L. R. Squire, and D. G. Amaral, “Three cases of enduring memory impairment after bilateral damage limited to the hippocampal formation,” Journal of Neuroscience , vol. 16, no. 16, pp. 5233–5255, 1996. [Online]. Available: https: //www.jneurosci.org/content/16/16/5233

work page 1996
[8]

Memory for famous people in patients with unilateral temporal lobe epilepsy and excisions,

I. V . Viskontas, M. P. McAndrews, and M. Moscovitch, “Memory for famous people in patients with unilateral temporal lobe epilepsy and excisions,” Neuropsychology, vol. 16, no. 4, pp. 472–480, 2002

work page 2002

Show all 29 references

[9]

Remote spatial memory in an amnesic person with extensive bilateral hippocampal lesions,

R. S. Rosenbaum, S. Priselac, S. K ¨ohler, S. E. Black, F. Gao, L. Nadel, and M. Moscovitch, “Remote spatial memory in an amnesic person with extensive bilateral hippocampal lesions,” Nat. Neurosci., vol. 3, no. 10, pp. 1044–1048, Oct. 2000

2000
[10]

Long-term retrograde amnesia. . . the crucial role of the hippocampus,

L. Cipolotti, T. Shallice, D. Chan, N. Fox, R. Scahill, G. Harrison, J. Stevens, and P. Rudge, “Long-term retrograde amnesia. . . the crucial role of the hippocampus,” Neuropsychologia, vol. 39, no. 2, pp. 151–172, 2001. [Online]. Available: https://www.sciencedirect.com/ scie...

2001
[11]

Multiple trace theory of human memory: Computational, neuroimaging, and neuropsychological results,

L. Nadel, A. Samsonovich, L. Ryan, and M. Moscovitch, “Multiple trace theory of human memory: Computational, neuroimaging, and neuropsychological results,” Hippocampus, vol. 10, no. 4, pp. 352–368, 2000. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1002/1098...

doi
[12]

Neural networks and physical systems with emergent collective computational abilities

J. J. Hopfield, “Neural networks and physical systems with emergent collective computational abilities.” Proceedings of the National Academy of Sciences , vol. 79, no. 8, pp. 2554–2558, 1982. [Online]. Available: https://www.pnas.org/doi/abs/10.1073/pnas.79.8.2554

1982 doi
[13]

Learning andf relearning in boltzmann machines,

G. E. Hinton and T. J. Sejnowski, “Learning andf relearning in boltzmann machines,” in Parallel Distributed Processing, Volume 1: Explorations in the Microstructure of Cognition: Foundations . The MIT Press, 07 1986. [Online]. Available: https://doi.org/10.7551/ mitpress/5236.003.0011

1986
[14]

A new frontier for hopfield networks,

D. Krotov, “A new frontier for hopfield networks,” Nature Reviews Physics, vol. 5, no. 7, p. 366–367, May 2023. [Online]. Available: http://dx.doi.org/10.1038/s42254-023-00595-y

2023 doi
[15]

Overparameterized neural networks can implement associative memory,

A. Radhakrishnan, M. Belkin, and C. Uhler, “Overparameterized neural networks can implement associative memory,” CoRR, vol. abs/1909.12362, 2019. [Online]. Available: http://arxiv.org/abs/1909. 12362

1909 arXiv
[16]

Learning and inference in the brain,

K. Friston, “Learning and inference in the brain,” Neural Networks , vol. 16, no. 9, pp. 1325–1352, 2003

2003
[17]

Associative memories via predictive coding,

T. Salvatori, Y . Song, Y . Hong, S. Frieder, L. Sha, Z. Xu, R. Bogacz, and T. Lukasiewicz, “Associative memories via predictive coding,”
[18]

Place cells may simply be memory cells: Memory compression leads to spatial tuning and history dependence,

M. K. Benna and S. Fusi, “Place cells may simply be memory cells: Memory compression leads to spatial tuning and history dependence,” Proceedings of the National Academy of Sciences , vol. 118, no. 51, p. e2018422118, 2021

2021
[19]

Entorhinal mismatch: A model of self- supervised learning in the hippocampus,

D. Santos-Pata, A. F. Amil, I. G. Raikov, C. Renn ´o-Costa, A. Mura, I. Soltesz, and P. F. Verschure, “Entorhinal mismatch: A model of self- supervised learning in the hippocampus,” Iscience, vol. 24, no. 4, 2021

2021
[20]

Prediction and memory: A predictive coding account,

H. C. Barron, R. Auksztulewicz, and K. Friston, “Prediction and memory: A predictive coding account,” Progress in neurobiology , vol. 192, p. 101821, 2020

2020
[21]

Emergence of simple-cell receptive field properties by learning a sparse code for natural images,

B. A. Olshausen and D. J. Field, “Emergence of simple-cell receptive field properties by learning a sparse code for natural images,” Nature, vol. 381, no. 6583, pp. 607–609, 1996

1996
[22]

Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects,

R. P. Rao and D. H. Ballard, “Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects,” Nature neuroscience, vol. 2, no. 1, pp. 79–87, 1999

1999
[23]

Attention, uncertainty, and free-energy,

H. Feldman and K. Friston, “Attention, uncertainty, and free-energy,” Frontiers in Human Neuroscience , vol. 4, 2010. [Online]. Avail- able: https://www.frontiersin.org/journals/human-neuroscience/articles/ 10.3389/fnhum.2010.00215

2010 arXiv
[24]

A stable, fast, and fully automatic learning algorithm for predictive coding networks,

T. Salvatori, Y . Song, Y . Yordanov, B. Millidge, Z. Xu, L. Sha, C. Emde, R. Bogacz, and T. Lukasiewicz, “A stable, fast, and fully automatic learning algorithm for predictive coding networks,” 2024. [Online]. Available: https://arxiv.org/abs/2212.00720

2024 arXiv
[25]

Implementing predictive processing and active inference: Preliminary steps and results,

B. Millidge, “Implementing predictive processing and active inference: Preliminary steps and results,” Mar. 2019. [Online]. Available: http://dx.doi.org/10.31234/osf.io/4hb58

2019 doi
[26]

Bad minima of predictive coding energy functions,

S. Frieder, L. Pinchetti, and T. Lukasiewicz, “Bad minima of predictive coding energy functions,” in The Second Tiny Papers Track at ICLR 2024, 2024. [Online]. Available: https://openreview.net/forum?id= 5Vc1ACZ8eC

2024
[27]

Benchmarking predictive coding networks – made simple,

L. Pinchetti, C. Qi, O. Lokshyn, G. Olivers, C. Emde, M. Tang, A. M’Charrak, S. Frieder, B. Menzat, R. Bogacz, T. Lukasiewicz, and T. Salvatori, “Benchmarking predictive coding networks – made simple,” 2024. [Online]. Available: https://arxiv.org/abs/2407.01163

2024 arXiv
[28]

Associative memories via predictive coding,

T. Salvatori, Y . Song, Y . Hong, L. Sha, S. Frieder, Z. Xu, R. Bogacz, and T. Lukasiewicz, “Associative memories via predictive coding,”Advances in Neural Information Processing Systems , vol. 34, pp. 3874–3886, 2021

2021
[2021]

Available: https://arxiv.org/abs/2109.08063

[Online]. Available: https://arxiv.org/abs/2109.08063

arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

A generative model of memory construction and consolidation,

E. Spens and N. Burgess, “A generative model of memory construction and consolidation,” Nature Human Behaviour , pp. 1–18, 2024

work page 2024

[2] [2]

Off-line replay maintains declarative memories in a model of hippocampal-neocortical interactions,

S. K ´ali and P. Dayan, “Off-line replay maintains declarative memories in a model of hippocampal-neocortical interactions,” Nature Neuroscience, vol. 7, no. 3, pp. 286–294, 2004

work page 2004

[3] [3]

What learning systems do intelligent agents need? complementary learning systems theory updated,

D. Kumaran, D. Hassabis, and J. L. McClelland, “What learning systems do intelligent agents need? complementary learning systems theory updated,” Trends in cognitive sciences , vol. 20, no. 7, pp. 512–534, 2016

work page 2016

[4] [4]

Retrograde amnesia and memory consolidation: a neurobiological perspective,

L. R. Squire and P. Alvarez, “Retrograde amnesia and memory consolidation: a neurobiological perspective,” Current Opinion in Neurobiology, vol. 5, no. 2, pp. 169–177, 1995. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0959438895800239

arXiv 1995

[5] [5]

Anterograde and retrograde amnesia in rats with dorsal hippocampal or dorsomedial thalamic lesions,

G. Winocur, “Anterograde and retrograde amnesia in rats with dorsal hippocampal or dorsomedial thalamic lesions,” Behavioural Brain Research, vol. 38, no. 2, pp. 145–154, 1990. [Online]. Available: https://www.sciencedirect.com/science/article/pii/0166432890900124

work page arXiv 1990

[6] [6]

The primate hippocampal formation: Evidence for a time-limited role in memory storage,

S. M. Zola-Morgan and L. R. Squire, “The primate hippocampal formation: Evidence for a time-limited role in memory storage,” Science, vol. 250, no. 4978, pp. 288–290, 1990. [Online]. Available: https://www.science.org/doi/abs/10.1126/science.2218534

work page doi:10.1126/science.2218534 1990

[7] [7]

Three cases of enduring memory impairment after bilateral damage limited to the hippocampal formation,

N. L. Rempel-Clower, S. M. Zola, L. R. Squire, and D. G. Amaral, “Three cases of enduring memory impairment after bilateral damage limited to the hippocampal formation,” Journal of Neuroscience , vol. 16, no. 16, pp. 5233–5255, 1996. [Online]. Available: https: //www.jneurosci.org/content/16/16/5233

work page 1996

[8] [8]

Memory for famous people in patients with unilateral temporal lobe epilepsy and excisions,

I. V . Viskontas, M. P. McAndrews, and M. Moscovitch, “Memory for famous people in patients with unilateral temporal lobe epilepsy and excisions,” Neuropsychology, vol. 16, no. 4, pp. 472–480, 2002

work page 2002