Pith. sign in

REVIEW 5 major objections 5 minor 51 references

Computing with Canonical Microcircuits

T0 review · 5 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper proposes that canonical microcircuit motifs from the neocortex, implemented as learnable 8-dimensional neural ODEs, can classify images competitively and interpretably using far fewer parameters than conventional deep networks.

desk verdict A genuinely new CMC-as-nODE combination undermined by internally inconsistent accuracy numbers; the central claim can't be trusted as submitted. read the letter →

arxiv 2508.06501 v1 pith:H4B57T26 submitted 2025-07-25 q-bio.NC cs.AIcs.NE

PK Douglas This is my paper · ORCID
classification q-bio.NCcs.AIcs.NE
keywords canonicalmicrocircuitsneuralODErecurrentdynamicsneuromorphiccomputingparameterefficiencyvisualhierarchyMNISTphasespaceanalysis
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 tries to establish that the stereotyped four-population circuits of the neocortex, translated into learnable neural ordinary differential equations, are viable building blocks for computer vision. It reports that even a single canonical microcircuit node reaches 97.8 percent accuracy on MNIST, and that stacking one to five such nodes improves performance on more demanding image benchmarks while using only about 150,000 parameters, orders of magnitude fewer than conventional convolutional networks. If true, this would mean that biologically derived recurrent dynamics can replace much larger learned feature extractors and also supply interpretable, class-specific trajectories in state space. The practical payoff claimed is a path toward parameter-efficient, interpretable neuromorphic models grounded in real cortical circuit motifs.

What carries the argument

The load-bearing object is the canonical microcircuit (CMC) neural ODE: a four-population recurrent dynamical system in which spiny stellate cells, inhibitory interneurons, superficial pyramidal cells, and deep pyramidal cells each obey a second-order differential equation derived from a standard predictive-coding microcircuit model. Written as first-order equations, each node is 8-dimensional; its connectivity strengths ($\gamma$) and time constants ($\kappa$) are learnable, and the whole system is integrated with a fixed solver over a short time window as the hidden-state evolution of a neural ODE. This machinery converts a neuroscience circuit model into a trainable module, and repeated modules with inter-node connectivity form the hierarchy used for vision tasks.

What would settle it

Train the same architecture with the canonical microcircuit ODE replaced by a plain feedforward or linear layer of comparable parameter count, keeping the retinal preprocessing, state-initialization projection, readout, and solver schedule fixed; if MNIST accuracy stays near 97.8 percent, the microcircuit dynamics are not the source of the claimed capability.

Watch

Extended reading notes

Core claim

The central claim is that canonical microcircuits—stereotyped laminar patterns of spiny stellate, inhibitory, and pyramidal neurons—can be operationalized as an 8-dimensional neural ODE system and trained end-to-end on raw image classification. Each cortical area is modeled by this ODE, with learnable synaptic strengths and time constants, and multiple areas are connected by learnable forward and backward matrices that mirror the V1-to-V5 visual hierarchy. The paper reports competitive MNIST and CIFAR-10 accuracy with roughly 150,000 parameters for a five-region model, and shows that different digit classes converge to distinct phase-space attractors during inference. It concludes that biologically plausible circuit dynamics, not just brain-inspired layer wiring, can deliver both efficiency and interpretability.

Load-bearing premise

The paper attributes its accuracy and parameter efficiency to the microcircuit, but every model also contains a learnable retina-like preprocessing layer, a learned projection that sets the initial state, a learned readout, and hand-selected solver settings, and no control replaces the microcircuit ODE with an ordinary layer to show those parts are not responsible for the gains.

Editorial extensions

If this is right

  • A single CMC node can serve as a parameter-efficient feature extractor for MNIST, with a five-node hierarchy reaching about 99 percent test accuracy at roughly 150,000 parameters.
  • Scaling from one to five nodes consistently improves accuracy on CIFAR-10, suggesting that deeper hierarchies of the same recurrent motif extract progressively more abstract features.
  • Removing inter-regional recurrence yields worse generalization and rougher loss surfaces than the recurrent version, so recurrent feedback across nodes is doing real work.
  • Class-specific attractors in phase space give a built-in interpretability channel: the dynamics themselves, not a post-hoc explanation, distinguish input classes.

Reading between the lines

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

  • The paper leaves open whether the microcircuit dynamics or the learnable preprocessing and readout surrounding them explain the gains; a matched control that replaces the CMC ODE with a plain feedforward layer would isolate the contribution.
  • If the recurrent ODE is doing the heavy lifting, any sufficiently expressive recurrent ODE with comparable weight sharing might reproduce the parameter efficiency, meaning the biological detail may matter less than recurrence itself.
  • The observed class-specific attractors suggest a testable extension the author does not pursue: using distance to each class attractor as a confidence score for out-of-distribution detection or few-shot classification.
  • Because the solver uses a fixed short time window, the model is effectively a depth-unrolled RNN; comparing against a discrete RNN with the same integration schedule would quantify what continuous-time dynamics add.
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.

Referee Report

5 major / 5 minor

Summary. The paper proposes a neural ODE implementation of canonical microcircuits (CMCs) for image classification. The architecture combines a learnable retinal preprocessing front-end, state-initialization projection, one or more CMC nodes with eight-dimensional dynamics derived from Bastos et al., learnable inter-node connectivity, and a learned readout. The authors report MNIST and CIFAR-10 results for one to five nodes, claim orders-of-magnitude parameter efficiency relative to conventional deep networks, and present phase-space trajectories as evidence of interpretable class-specific dynamics.

Significance. If the empirical claims held, this would demonstrate that a biologically derived recurrent ODE module with roughly 150K parameters can match much larger convolutional or transformer models on standard vision benchmarks while offering interpretable dynamics. The paper's strengths include the use of established CMC dynamics, a clear biological motivation, and an exploratory ablation study. However, the headline accuracy figures are internally inconsistent across sections, no matched control isolates the effect of the CMC ODE, and the parameter-efficiency comparison lacks a quantitative baseline, so the significance cannot be assessed from the manuscript as submitted.

major comments (5)
  1. [Abstract; Sections 5.2, 6.2, 6.3; Figure 4] The abstract's headline claim that a single CMC node achieves 97.8% accuracy on MNIST is not reported anywhere in the experimental sections: Section 6.2 reports 86.4% for one node and 95.2% for four nodes, Section 6.3 reports 86.5% and 99.2% for the same two configurations, Section 5.2 reports 99.8% for a five-region model, and the learning curves in Figure 4 appear to plateau near 93-94% for all node counts. These numbers cannot all describe the same experimental setup, and the manuscript provides no reconciliation or confidence intervals, so the central empirical claim is unverifiable.
  2. [Sections 4.1, 5.2, 5.3] Every trained model contains a learnable retinal preprocessing front-end, a learned state-initialization projection, a learned readout, and learnable inter-node connectivity matrices, yet the paper never compares against a matched control in which the CMC ODE is replaced by an ordinary feedforward or linear map with the same surrounding modules. Without such a control, the accuracy and parameter-efficiency gains cannot be attributed to the microcircuit dynamics, which is the central claim of the paper.
  3. [Section 5.3; Section 6.3; Figure 1] The claimed 'orders of magnitude fewer parameters' comparison is not substantiated: no table lists baseline architectures with their parameter counts and accuracies, and the reported parameter count for the model is inconsistent (Section 5.3 says a 5-region model has approximately 150,000 parameters, while Section 6.3 attributes 150K parameters to the 4-node model). This prevents any quantitative assessment of the parameter-efficiency claim.
  4. [Supplementary Material, Section A; Section 7] The supplementary ablation section states that 'the original model was outperformed by each of the ablation tests for a single V1 node,' yet Section 7 summarizes the ablations as showing that 'both architecture and neural population dynamics contribute substantially to overall performance.' These statements are contradictory, and the single-node result in particular undermines the claim that each cell type is essential to the CMC's performance.
  5. [Sections 5.1 and 6.1] The numerical solver specification is internally inconsistent: Section 5.1 says the integration time span is [0,1] with 6 time points and a step size of 0.15, while Section 6.1 says Euler integration with a step size of 0.05 was used. The relation between '6 time points' and the time span is not defined, so the experiments cannot be reproduced exactly from the description.
minor comments (5)
  1. [Section 6.2] The text says 'Figure 2 shows test accuracy across datasets' but Figure 2 depicts laminar structure; the accuracy curves appear in Figure 4.
  2. [Section 6.4] The text refers to 'Figure 4' when describing tracked voltage evolution during training, but Figure 4 contains learning curves; the phase portraits are in Figure 5 and Figure 6, and the voltage dynamics are in Figure 3. The cross-references need correction.
  3. [Section 3.0.1 and Section 3.1] The paper refers to 'Section 3.1' for the ODE dynamics, but the equations appear under the malformed subsection heading '3.0.1 Neural Dynamics'; the section numbering should be repaired.
  4. [Figure 3] The lower portion of Figure 3 appears to reproduce the abstract of Chen et al. (2018) and a 'Model Performance' bar chart without visible attribution or discussion; if this is an unintended insertion, it should be removed, and if intentional, proper permission and citation are required.
  5. [Throughout] There are several typographical errors, including 'acheived' in Section 6.3 and 'feed foward' in the supplementary material; a careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CMC dynamics are imported from external prior work and tested against external benchmarks, so the main claims are not equivalent to their inputs by construction.

full rationale

The derivation chain is: adopt Bastos et al. [6] canonical microcircuit ODEs (Eqs. 1–4); embed them in a trainable neural-ODE pipeline with retinal preprocessing, state initialization, and readout; train on MNIST and CIFAR-10; measure test accuracy. No claimed result is defined in terms of the quantity it purports to predict. The ODE form is explicitly attributed to prior external work ('we implement the standard instantiation proposed by [6]'), not to an ansatz smuggled in via self-citation. The only self-citation ([15], Kriegeskorte and Douglas) supports the framing remark that cross-pollination between neuroscience and AI 'continues to generate excitement'; it is not load-bearing for any experimental claim. The 'class-specific attractors' are post-hoc visualizations, not predictions used to derive accuracy. The paper does not run a matched feedforward control, and the abstract's 97.8% single-node MNIST figure is not reproduced in Sections 6.2 and 6.3 (86.4%/86.5% for one node; 95.2%/99.2% for four), but these are correctness and verifiability concerns, not circularity: no equation in the paper reduces one claim to another. Under the hard rule that circularity requires exhibiting a specific reduction (Eq. X = Eq. Y by construction, or a fitted parameter renamed as prediction), none is present.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The central result is an empirical machine-learning claim, so learnable parameters are expected rather than hidden assumptions. The deeper burden is that the dynamics, the discretization, and the per-area aggregation are all taken as faithful enough to preserve the biological inductive bias, while the parameter-efficiency count does not separate the CMC internals from the surrounding trainable projections. No new physical entities are postulated; all biological vocabulary is imported from prior work.

free parameters (7)
  • Intra-node connectivity gamma (Eqs. 1-4)
    All coupling weights are learned by Adam on the target dataset; no learned values or distributions are reported.
  • Synaptic time constants kappa_e, kappa_i
    Stated as learnable in Section 5.1; no fitted values are given.
  • Inter-node connectivity matrices
    Learnable coupling between CMC nodes (Section 4.3) drives hierarchical performance; no values reported.
  • Input projection and state initialization network
    A linear layer maps retinal features to initial ODE states (Section 4.1); its parameters contribute to the total count but are not separated.
  • Output projection network
    Maps final states to class logits (Section 5.2); contributes to parameter count and performance.
  • Retinal preprocessing kernels and gain gates
    Learnable convolutional center-surround, ON/OFF, lateral inhibition, and sigmoid-gated gain (Section 4.1); not accounted for in the efficiency discussion.
  • Solver hyperparameters (step size, time span)
    Chosen by hand for stability; stated as 0.15 in Section 5.1 and 0.05 in Section 6.1, with no sensitivity analysis.
assumptions (4)
  • domain assumption The Bastos et al. canonical microcircuit equations (Eqs. 1-4) capture the computationally relevant dynamics of cortical columns.
    Adopted unchanged from Ref [6]; Section 3.0.1 says the goal is not to invent a new dynamical form, so all biological credibility is inherited from that prior model.
  • domain assumption A coarse Euler or midpoint discretization over the interval [0,1] with normalization preserves the ODE's computational content.
    Solver choices in Sections 5.1 and 6.1 are for speed and stability, but no convergence study is given, and the two sections state different step sizes.
  • domain assumption One CMC node per cortical area with mean-field inter-node coupling is a sufficient abstraction of the visual hierarchy.
    Section 4.3 models inter-node influence as global mean activity, a simplification that may discard the cell-type-specific connectivity the single-node equations are built on.
  • ad hoc to paper Visualizing class-specific phase trajectories is evidence of interpretable emergent behavior.
    Section 6.4 offers this interpretation without quantitative measures of separability or comparison to randomized dynamics.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Computing with Canonical Microcircuits." pith.science (2026). https://pith.science/paper/H4B57T26

@misc{pith2026250806501,
  author       = {Pith},
  title        = {Pith review of: Computing with Canonical Microcircuits},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H4B57T26}},
  note         = {Machine review of arXiv:2508.06501}
}
read the original abstract

The human brain represents the only known example of general intelligence that naturally aligns with human values. On a mere 20-watt power budget, the brain achieves robust learning and adaptive decision-making in ways that continue to elude advanced AI systems. Inspired by the brain, we present a computational architecture based on canonical microcircuits (CMCs) - stereotyped patterns of neurons found ubiquitously throughout the cortex. We implement these circuits as neural ODEs comprising spiny stellate, inhibitory, and pyramidal neurons, forming an 8-dimensional dynamical system with biologically plausible recurrent connections. Our experiments show that even a single CMC node achieves 97.8 percent accuracy on MNIST, while hierarchical configurations - with learnable inter-regional connectivity and recurrent connections - yield improved performance on more complex image benchmarks. Notably, our approach achieves competitive results using substantially fewer parameters than conventional deep learning models. Phase space analysis revealed distinct dynamical trajectories for different input classes, highlighting interpretable, emergent behaviors observed in biological systems. These findings suggest that neuromorphic computing approaches can improve both efficiency and interpretability in artificial neural networks, offering new directions for parameter-efficient architectures grounded in the computational principles of the human brain.

Figures

Figures reproduced from arXiv: 2508.06501 by the authors.

Figure 1
Figure 1. Our brain-inspired model achieves competi￾tive performance with orders of magnitude fewer param￾eters than conventional networks. Canonical microcircuits (CMCs) capture the population-level dynamics of functionally dis￾tinct neurons within a cortical column [6][7]. Each "neuron" in these models represents the collective behavior of a subpopulation of neu￾rons sharing the same cell type, cortical layer, microcolumn m… view at source ↗
Figure 2
Figure 2. Laminar structure and repeated cytoarchitecture in the neocortex. Left: Cortical parcellation [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Single CMC node. Voltage dynamics (left) over time, and 3D loss landscape for a single [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Top row: Accuracy and learning curves for CMC models with 1–5 regions on MNIST (top [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Effect of hierarchical depth on optimization and parameter structure in the CMC-nODE [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Phase portraits of superficial pyramidal cell activity for digits 1 (blue) and 3 (orange) across [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: This figure depicts the full Canonical Microcircuit (CMC) model (top left, S7.1), alongside [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Neural ODE formulations for the original canonical microcircuit (CMC), and each neural [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: Loss landscapes for feedforward-only CMC models on CIFAR-10. Top row: cortical [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Voltage dynamics of neuronal populations in the V1 CMC node following cell-type [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]
Figure 11
Figure 11. Figure 11: Hierarchical Ablation Analysis on MNIST. Top-left: Test accuracy as a function of [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Hierarchical Ablation Analysis on CIFAR10. Top-left: Test accuracy as a function of [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 29 canonical work pages

  1. [1]

    Excitatory and Inhibitory Interactions in Localized Populations of Model Neurons,

    H. R. Wilson and J. D. Cowan, “Excitatory and Inhibitory Interactions in Localized Populations of Model Neurons,” Biophysical Journal, vol. 12, no. 1, pp. 1–24, Jan. 1972, doi: 10.1016/S0006- 3495(72)86068-5

    doi:10.1016/s0006- 1972

  2. [2]

    NEURONAL CIRCUITS OF THE NEOCOR- TEX,

    R. J. Douglas and K. A. C. Martin, “NEURONAL CIRCUITS OF THE NEOCOR- TEX,” Annu. Rev. Neurosci., vol. 27, no. 1, pp. 419–451, Jul. 2004, doi: 10.1146/an- nurev.neuro.27.070203.144152

    arXiv 2004

  3. [3]

    The columnar organization of the neocortex,

    V . Mountcastle, “The columnar organization of the neocortex,” Brain, vol. 120, no. 4, pp. 701–722, Apr. 1997, doi: 10.1093/brain/120.4.701

    doi:10.1093/brain/120.4.701 1997

  4. [4]

    The neocortical circuit: themes and variations,

    K. D. Harris and G. M. G. Shepherd, “The neocortical circuit: themes and variations,” Nat Neurosci, vol. 18, no. 2, pp. 170–181, Feb. 2015, doi: 10.1038/nn.3917

    doi:10.1038/nn.3917 2015

  5. [5]

    Dynamics of neocortical networks: connectivity beyond the canonical microcircuit,

    H. J. Luhmann, “Dynamics of neocortical networks: connectivity beyond the canonical microcircuit,” Pflugers Arch - Eur J Physiol, vol. 475, no. 9, pp. 1027–1033, Sep. 2023, doi: 10.1007/s00424-023-02830-y

    work page doi:10.1007/s00424-023-02830-y 2023

  6. [6]

    Canonical Microcircuits for Predictive Coding,

    A. M. Bastos, W. M. Usrey, R. A. Adams, G. R. Mangun, P. Fries, and K. J. Friston, “Canonical Microcircuits for Predictive Coding,” Neuron, vol. 76, no. 4, pp. 695–711, Nov. 2012, doi: 10.1016/j.neuron.2012.10.038

    doi:10.1016/j.neuron.2012.10.038 2012

  7. [7]

    Neural masses and fields in dynamic causal modeling,

    R. Moran, D. A. Pinotsis, and K. Friston, “Neural masses and fields in dynamic causal modeling,” Front. Comput. Neurosci., vol. 7, 2013, doi: 10.3389/fncom.2013.00057

    arXiv 2013

  8. [8]

    A column-like organization for ocular dominance in mouse visual cortex,

    P. M. Goltstein, D. Laubender, T. Bonhoeffer, and M. Hübener, “A column-like organization for ocular dominance in mouse visual cortex,” Nat Commun, vol. 16, no. 1, p. 1926, Feb. 2025, doi: 10.1038/s41467-025-56780-3

    work page doi:10.1038/s41467-025-56780-3 1926

Show all 51 references
  1. [9]

    Layer-specificity in the effects of attention and working memory on activity in primary visual cortex,

    T. Van Kerkoerle, M. W. Self, and P. R. Roelfsema, “Layer-specificity in the effects of attention and working memory on activity in primary visual cortex,” Nat Commun, vol. 8, no. 1, p. 13804, Jan. 2017, doi: 10.1038/ncomms13804

    2017 doi

  2. [10]

    Laminar compartmentalization of attention modulation in area V4 aligns with the demands of visual processing hierarchy in the cortex,

    X. Wang, A. S. Nandy, and M. P. Jadi, “Laminar compartmentalization of attention modulation in area V4 aligns with the demands of visual processing hierarchy in the cortex,” Sci Rep, vol. 13, no. 1, p. 19558, Nov. 2023, doi: 10.1038/s41598-023-46722-8

    2023 doi

  3. [11]

    Cracking the Function of Layers in the Sensory Cortex,

    H. Adesnik and A. Naka, “Cracking the Function of Layers in the Sensory Cortex,” Neuron, vol. 100, no. 5, pp. 1028–1043, Dec. 2018, doi: 10.1016/j.neuron.2018.10.032

    2018 doi

  4. [12]

    Neural effects of propofol-induced unconsciousness and its reversal using thalamic stimulation,

    A. M. Bastos et al., “Neural effects of propofol-induced unconsciousness and its reversal using thalamic stimulation,” eLife, vol. 10, p. e60824, Apr. 2021, doi: 10.7554/eLife.60824

    2021 doi

  5. [13]

    A logical calculus of the ideas immanent in nervous activity,

    W. S. McCulloch and W. Pitts, “A logical calculus of the ideas immanent in nervous activity,” Bulletin of Mathematical Biophysics, vol. 5, no. 4, pp. 115–133, Dec. 1943, doi: 10.1007/BF02478259. 10

    1943 doi

  6. [14]

    Catalyzing next-generation Artificial Intelligence through NeuroAI,

    A. Zador et al., “Catalyzing next-generation Artificial Intelligence through NeuroAI,” Nat Commun, vol. 14, no. 1, p. 1597, Mar. 2023, doi: 10.1038/s41467-023-37180-x

    2023 doi

  7. [15]

    Cognitive computational neuroscience,

    N. Kriegeskorte and P. K. Douglas, “Cognitive computational neuroscience,” Nat Neurosci, vol. 21, no. 9, pp. 1148–1160, Sep. 2018, doi: 10.1038/s41593-018-0210-5

    2018 doi

  8. [16]

    Towards Understanding Learning Representations: To What Extent Do Different Neural Networks Learn the Same Representation,

    L. Wang et al., “Towards Understanding Learning Representations: To What Extent Do Different Neural Networks Learn the Same Representation,” Nov. 28, 2018, arXiv: arXiv:1810.11750. Accessed: May 19, 2022. [Online]. Available: http://arxiv.org/abs/1810.11750

    2018 arXiv

  9. [17]

    A functional microcircuit for cat visual cortex.,

    R. J. Douglas and K. A. Martin, “A functional microcircuit for cat visual cortex.,” The Journal of Physiology, vol. 440, no. 1, pp. 735–769, Aug. 1991, doi: 10.1113/jphysiol.1991.sp018733

    1991 doi

  10. [18]

    Efficient codes and balanced networks,

    S. Denève and C. K. Machens, “Efficient codes and balanced networks,” Nat Neurosci, vol. 19, no. 3, pp. 375–382, Mar. 2016, doi: 10.1038/nn.4243

    2016 doi

  11. [19]

    Inputs to prefrontal cortex support visual recognition in the aging brain,

    J. R. Gilbert and R. J. Moran, “Inputs to prefrontal cortex support visual recognition in the aging brain,” Sci Rep, vol. 6, no. 1, p. 31943, Aug. 2016, doi: 10.1038/srep31943

    2016 doi

  12. [20]

    A model of individualized canonical microcircuits supporting cognitive operations,

    T. Kunze, A. D. H. Peterson, J. Haueisen, and T. R. Knösche, “A model of individualized canonical microcircuits supporting cognitive operations,” PLoS ONE, vol. 12, no. 12, p. e0188003, Dec. 2017, doi: 10.1371/journal.pone.0188003

    2017 doi

  13. [21]

    A mechanistic cortical microcircuit of attention for amplification, normalization and suppression,

    F. Beuth and F. H. Hamker, “A mechanistic cortical microcircuit of attention for amplification, normalization and suppression,” Vision Research, vol. 116, pp. 241–257, Nov. 2015, doi: 10.1016/j.visres.2015.04.004

    2015 doi

  14. [22]

    Neural Ordinary Differential Equations,

    R. T. Q. Chen, Y . Rubanova, J. Bettencourt, and D. K. Duvenaud, “Neural Ordinary Differential Equations,” in Advances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., Curran Associates, Inc., 2018

    2018

  15. [23]

    A Proposal on Machine Learning via Dynamical Systems,

    E. Weinan, “A Proposal on Machine Learning via Dynamical Systems,” Commun. Math. Stat., vol. 5, no. 1, pp. 1–11, Mar. 2017, doi: 10.1007/s40304-017-0103-z

    2017 doi

  16. [24]

    Stable Architectures for Deep Neural Networks,

    E. Haber and L. Ruthotto, “Stable Architectures for Deep Neural Networks,” 2017, doi: 10.48550/ARXIV .1705.03341

    2017 doi arXiv

  17. [25]

    Neural Controlled Differential Equations for Irregular Time Series,

    P. Kidger, J. Morrill, J. Foster, and T. Lyons, “Neural Controlled Differential Equations for Irregular Time Series,” 2020, arXiv. doi: 10.48550/ARXIV .2005.08926

    2020 doi arXiv

  18. [26]

    FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models,

    W. Grathwohl, R. T. Q. Chen, J. Bettencourt, I. Sutskever, and D. Duvenaud, “FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models,” 2018, arXiv. doi: 10.48550/ARXIV .1810.01367

    2018 doi arXiv

  19. [27]

    Latent ODEs for Irregularly-Sampled Time Series,

    Y . Rubanova, R. T. Q. Chen, and D. Duvenaud, “Latent ODEs for Irregularly-Sampled Time Series,” 2019, arXiv. doi: 10.48550/ARXIV .1907.03907

    2019 doi arXiv

  20. [28]

    Augmented Neural ODEs,

    E. Dupont, A. Doucet, and Y . W. Teh, “Augmented Neural ODEs,” 2019, arXiv. doi: 10.48550/ARXIV .1904.01681. 11

    2019 doi arXiv

  21. [29]

    Scalable Gradients for Stochastic Differential Equations,

    X. Li, T.-K. L. Wong, R. T. Q. Chen, and D. Duvenaud, “Scalable Gradients for Stochastic Differential Equations,” 2020, arXiv. doi: 10.48550/ARXIV .2001.01328

    2020 doi arXiv

  22. [30]

    Graph Neural Ordinary Differential Equations,

    M. Poli, S. Massaroli, J. Park, A. Yamashita, H. Asama, and J. Park, “Graph Neural Ordinary Differential Equations,” 2019, arXiv. doi: 10.48550/ARXIV .1911.07532

    2019 doi arXiv

  23. [31]

    A Forward Learning Algorithm for Neural Memory Ordinary Differential Equations,

    X. Xu, H. Luo, Z. Yi, and H. Zhang, “A Forward Learning Algorithm for Neural Memory Ordinary Differential Equations,” Int. J. Neur. Syst., vol. 34, no. 09, p. 2450048, Sep. 2024, doi: 10.1142/S0129065724500485

    2024 doi

  24. [32]

    Functional specificity in the human brain: A window into the functional architecture of the mind,

    N. Kanwisher, “Functional specificity in the human brain: A window into the functional architecture of the mind,” Proc. Natl. Acad. Sci. U.S.A., vol. 107, no. 25, pp. 11163–11170, Jun. 2010, doi: 10.1073/pnas.1005062107

    2010 doi

  25. [33]

    Deep Neural Networks: A New Framework for Modeling Biological Vision and Brain Information Processing,

    N. Kriegeskorte, “Deep Neural Networks: A New Framework for Modeling Biological Vision and Brain Information Processing,” Annu. Rev. Vis. Sci., vol. 1, no. 1, pp. 417–446, Nov. 2015, doi: 10.1146/annurev-vision-082114-035447

    2015 doi

  26. [34]

    A deep learning framework for neuroscience,

    B. A. Richards et al., “A deep learning framework for neuroscience,” Nat Neurosci, vol. 22, no. 11, pp. 1761–1770, Nov. 2019, doi: 10.1038/s41593-019-0520-2

    2019 doi

  27. [35]

    Locus Coeruleus tracking of prediction errors optimises cognitive flexibility: An Active Inference model,

    A. C. Sales, K. J. Friston, M. W. Jones, A. E. Pickering, and R. J. Moran, “Locus Coeruleus tracking of prediction errors optimises cognitive flexibility: An Active Inference model,” PLoS Comput Biol, vol. 15, no. 1, p. e1006267, Jan. 2019, doi: 10.1371/journal.pcbi.1006267. 1...

    2019 doi

  28. [36]

    Guidelines: We followed all NeurIPS guidelines related to claims

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: We summarize the contributions clearly in the abstract and expand on them in Sections 1 and 3. Guidelines: We followed...

  29. [37]

    We created a specific section of the paper (Section 7) to directly address limitations related to the goals set forth in this paper

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] , Justification: YES. We created a specific section of the paper (Section 7) to directly address limitations related to the goals set forth in this paper. Guideline...

  30. [38]

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [NA]

  31. [39]

    We clearly describe our model architecture, training procedure, and experiments

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  32. [40]

    Guidelines: We followed all guidelines with respect to NeurIPS code

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: Yes, a link to our publicly...

  33. [41]

    All code is publicly accessible via the link provided

    Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: Yes, we clearly describe all ...

  34. [42]

    13 Guidelines: We followed all NeurIPS guidelines for reporting

    Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: Yes, our figures show error bars or intervals for ...

  35. [43]

    The code provided allows flexibility with respect to resources available, as described in the manuscript

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: As described in the doc...

  36. [44]

    Guidelines: Yes all research was conducted to conform, in every respect, with the NeurIPS Code of Ethics

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: Yes, we read the linked material to the code of ethics. Guidelines: Yes all res...

  37. [45]

    Justification: We consider this to be foundational research, and not readily tied to any particular societal impact

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [NA] . Justification: We consider this to be foundational research, and not readily tied to any particular societal impact. Gu...

  38. [46]

    Justification: No such risks Guidelines: Guidelines followed

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] . Justification: No such risks...

  39. [47]

    Guidelines: Guidelines followed

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer:[Yes] Justification: All works cited a...

  40. [48]

    Justification: No assets used or obtained

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [NA] . Justification: No assets used or obtained. Guidelines: Guidelines followed

  41. [49]

    Justification: Not used

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? 1...

  42. [50]

    Justification: Human data not used

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  43. [51]

    Answer: [NA]

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the ...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.