Dynamical Archetype Analysis: Autonomous Computation

Abel Sagodi; Il Memming Park

REVIEW 4 major objections 5 minor 1 cited by

Dynamical Archetype Analysis: Autonomous Computation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper proposes a trajectory-only dissimilarity that assigns a neural system to the closest dynamical archetype by balancing a learned coordinate transformation against its complexity.

desk verdict A promising trajectory-based dissimilarity for comparing neural dynamics, undercut by a false invariance claim and loose theory; worth refereeing after major revision. read the letter →

arxiv 2507.05505 v2 pith:K2HCT2PU submitted 2025-07-07 math.DS q-bio.NC

Abel Sagodi , Il Memming Park This is my paper

classification math.DSq-bio.NC MSC 37C1037C1537M0592B20

keywords dynamicalarchetypestopologicalconjugacyeffectivebehaviordissimilaritymeasurecontinuousattractorsrecurrentneuralnetworksODEworkingmemory

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

This paper argues that neural computations should be grouped by their effective behavior over finite time scales, not only by their asymptotic attractor topology, and proposes a concrete way to measure that: Dynamical Archetype Analysis (DAA). The method keeps a small library of idealized dynamical systems—a single fixed point, a bistable switch, a bounded line attractor, a ring attractor, and a limit cycle—and measures how far any observed trajectory data is from each archetype. The distance has two parts: how well a learned invertible coordinate transformation can make the archetype's trajectories match the target's, and how complicated that transformation is. The paper shows numerically that this two-part dissimilarity recognizes noisy, deformed, and perturbed ring attractors as ring-attractor-like, and that it assigns recurrent networks trained on an angular integration task to the ring attractor archetype, where prior trajectory-based dissimilarity measures were fragile. If the approach holds, it gives a data-driven vocabulary for saying what computation a neural system is performing without first estimating its vector field.

What carries the argument

The central object is the dissimilarity pair $(\bar{d}_1, c)$: $\bar{d}_1$ is the expected squared distance between target trajectories and archetype trajectories mapped through a learned diffeomorphism $\Phi_\theta$, and $c = d_{\mathrm{cxty}}(\Phi_\theta)$ is the integrated deviation of the diffeomorphism's Jacobian from the identity matrix. A Neural ODE—a trainable continuous-time map defined by integrating a neural-network vector field—parameterizes $\Phi_\theta$, so the map is invertible and can approximate the needed coordinate transformations; the complexity penalty $c$ prevents a wild deformation from trivially zeroing the trajectory error. A Gronwall-type bound shows that small vector-field perturbations imply small finite-time trajectory deviations, so the trajectory mismatch lower-bounds the perturbation needed, and the trade-off between perturbation size and deformation complexity is analyzed as a Pareto frontier.

What would settle it

Take the canonical ring attractor archetype, generate target trajectories, then generate a second target by rotating those trajectories by ninety degrees (or reflecting them) and rerun the proposed dissimilarity against the archetype. If the reported dissimilarity differs from the unrotated case—or the complexity penalty jumps because the diffeomorphism must encode the rotation—then the measure is not invariant under the isometries the paper says it should tolerate, and archetype assignment can change with the coordinate frame.

Watch

Extended reading notes

Core claim

The paper's central claim is that effective behavior of an autonomous dynamical system can be quantified by a two-part dissimilarity to an archetype: the mean squared trajectory error after the best-learned diffeomorphic mapping, plus the complexity of that mapping. On approximate ring attractors—systems whose continuous attractor has been broken into slow manifolds or fixed points—this dissimilarity grows slowly and the archetype assignment remains correct, whereas existing methods report fragility. The same assignment works for high-dimensional recurrent networks trained as internal compasses: the ring attractor archetype fits their low-dimensional trajectories across network sizes. The authors frame this as the start of a library-based vocabulary for neural computation, organizing archetypes by the dimension of their invariant manifold and allowing composition by Cartesian products.

Load-bearing premise

The target and archetype trajectories are assumed to live in a common coordinate frame, so rotations, reflections, permutations, and other symmetry transformations of the target are not free: the learned diffeomorphism must absorb them and the dissimilarity grows as a result.

Editorial extensions

If this is right

Observed trajectory data alone—without estimating the underlying vector field—becomes enough to assign a neural system to one archetype from a small library.
Approximate continuous attractors, including noisy, deformed, and perturbed ring attractors, would be recognized as ring-attractor-like instead of being misclassified, resolving a previously reported fragility.
High-dimensional recurrent networks can be compared with low-dimensional archetypes through the learned diffeomorphism, so the method applies where the state space is large but the computation lives on a low-dimensional manifold.
The archetype library can be organized by invariant-manifold dimension and composed via Cartesian products, giving a grammar for building higher-dimensional computations from interpretable parts.
The dissimilarity pair provides a Pareto-style trade-off: systems are effectively similar when a simple transformation makes their finite-time trajectories nearly coincide, and the frontier can guide how much deformation to allow.

Reading between the lines

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

Extension: the coordinate-frame sensitivity is the natural next pressure point—if a rotation- and reflection-invariant alignment step were added, the method could be applied across animals or sessions without pre-aligned latent spaces.
Extension: the Pareto trade-off between perturbation and diffeomorphism complexity suggests a scalar effective-behavior distance at a fixed complexity budget, turning archetype assignment into a model-selection problem.
Extension: the dissimilarity could be used as an objective to discover new archetypes from data rather than fixing the library a priori, or to test whether a proposed library is complete for a given family of tasks.
Extension: because the measure integrates only over $[0, T_{\max}]$, two systems with identical finite-time behavior but different asymptotic fates will look similar; choosing $T_{\max}$ is therefore an implicit modeling decision that should be reported with every assignment.

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

Short version: the paper has a good idea and a shaky wrapper. The idea is to compare dynamical systems by fitting a complexity-penalized diffeomorphism between archetype and target trajectories, then scoring the fit. That is a real gap: DSA's local linearization and SPE's normal-form fitting both miss approximate continuous attractors and high-dimensional RNN motifs. The Neural ODE estimator is concrete, and the experiments on deformed/perturbed ring attractors and a compass RNN show the method can group systems that DSA and SPE get wrong. The RNN result is the most convincing part.

The problems are mostly in the theory and the evaluation. The claimed invariance to rotations and affine transformations is false: Eq (5)'s trajectory distance can be zeroed by a rotation, but the complexity term Eq (6) penalizes every non-identity map, so a target in rotated coordinates gets a different (distance, complexity) pair than the same target in canonical coordinates. Since latent neural coordinates are arbitrary up to orthogonal and permutation transformations, this is load-bearing for the intended application. Also, Eq (5) does not vanish under a non-identity topological conjugacy as written; it only does if the conjugacy is the identity on the support of the sampling measure. The empirical estimator in Eq (23) correctly maps the initial condition back through Phi^{-1}, but the theoretical d1 in Eq (5) is mismatched, and the bounds in Eqs (11) and (22) inherit the same confusion.

The similarity scores are normalized per archetype across the target batch, so the best archetype for a target can change depending on what else is in the batch. No error bars, no code, no data. The compass RNN is trained with inputs, and the paper does not say how it is treated as autonomous.

None of this kills the core empirical idea. The method is novel, the synthetic results are plausible, and the RNN demonstration is suggestive. But the theoretical claims as stated are not defensible, and the rotation issue has to be fixed before I would trust the archetype assignments on real data. For someone comparing neural dynamics or designing similarity measures, the method is worth knowing about. A serious referee should see this; I would send it out with a clear request to address the invariance problem and the theory, and to release artifacts. I would not cite it in its current form.

Referee Report

4 major / 5 minor

Summary. This paper introduces Dynamical Archetype Analysis (DAA), a framework for comparing autonomous dynamical systems by their finite-time, trajectory-based "effective behavior." The core dissimilarity is a pair: (i) how well a source system's trajectories can be mapped onto a target system's trajectories by a diffeomorphism, and (ii) the complexity of that diffeomorphism measured by its deviation from the identity. The authors define a library of archetypes (single fixed point, bistable, bounded line attractor, ring attractor, limit cycle), estimate the dissimilarity from trajectories with a Neural ODE, and report experiments on deformed and perturbed ring attractors, nonlinear oscillators, and a task-trained compass RNN. They claim that the resulting archetype assignment overcomes failures of DSA and SPE on approximate continuous attractors and high-dimensional recurrent neural networks.

Significance. If the method is made rigorous, it would provide a practical, vector-field-free way to compare neural dynamics against interpretable dynamical motifs, and the explicit two-axis dissimilarity (trajectory distance plus deformation complexity) is a useful conceptual contribution. The archetype library is defined a priori, so the approach does not fit archetypes to data in a circular way, and the experiments cover several target classes including an RNN compass network. The paper does not supply machine-checked proofs or a code repository, and the theoretical bounds currently do not support the invariance claims; nevertheless the core idea is promising and the falsifiable predictions (ring-like targets map to the ring archetype, oscillator targets to the limit cycle archetype) are clearly stated.

major comments (4)

[§3.1, Eq. (5)] The statement that bar{d}_1 is zero when the source and target are topologically conjugate on the support of mu is not correct for the loss as written. If Phi is a conjugacy, then Phi(phi_g^t(x)) = phi_f^t(Phi(x)), whereas Eq. (5) compares phi_f^t(x) with Phi(phi_g^t(x)); these two trajectories agree only when Phi(x)=x for mu-almost every x. Thus the loss is zero only for diffeomorphisms that are the identity on the sampled region, not for non-identity conjugacies. Consequently, the bound in Eq. (11), which transfers the perturbation estimate Eq. (8) to bar{d}_1, does not follow from the preceding argument; the loss needs to be defined on matched initial conditions, for example bar{d}_1 = E_{xsimmu} int || phi_f^t(Phi(x)) − Phi(phi_g^t(x)) || dt, and a separate Gronwall estimate is needed for this quantity.
[§3.1, Eq. (6), and Supp. J.1.1–J.1.2] The claimed invariance to isometric scaling, rotations, and affine transformations is not realized by the dissimilarity that is actually computed. For a target that is a rotated copy of the archetype, the diffeomorphism Phi = R^{\-1} makes the trajectory term in Eq. (5) vanish, but the complexity penalty in Eq. (6) is int || R^{\-1} − I || dx, which is positive and increases with the rotation angle. Because the archetype assignment combines the similarity and simplicity scores (Supp. J.1.1–J.1.2), a target presented in rotated, reflected, or permuted coordinates receives a different dissimilarity pair than the same target in canonical coordinates. Since neural latent coordinates are arbitrary up to orthogonal and permutation transformations, this is a load-bearing issue for the application described in §3.1; the method must either optimize over the transformation equivalence class before measuring complexity or define complexity on the quotient by those transformations.
[Supp. C, Eq. (15)] The first-order computation contains a sign error. From y = x + epsilon Phitilde(x), equating dot{y} = f(x) + epsilon Delta(x) + epsilon nablaPhitilde(x) f(x) + O(epsilon^2) with dot{y} = f(y) = f(x) + epsilon nabla f(x) Phitilde(x) + O(epsilon^2) gives epsilon Delta = epsilon(nabla f)Phitilde − epsilon(nablaPhitilde) f, which is the negative of Eq. (15). The pointwise bound in Eq. (16) is unchanged in magnitude, but the displayed derivation should be corrected, and the role of Eq. (15) in the bound Eq. (22) should be rechecked.
[§3.1, Eq. (22)] The claimed upper bound in Eq. (22) is not established by the preceding inequalities. Eq. (8) bounds bar{d}_0(f, f+Delta), a comparison in a fixed coordinate frame with identical initial conditions, whereas Eq. (22) concerns bar{d}_1(f, g; c), which involves a diffeomorphism and trajectories evaluated at points related by that diffeomorphism. In addition, Delta_1 in Eqs. (9)–(10) is defined as the solution of a constrained optimization problem, and no proof is given that this solution satisfies the Lipschitz and boundedness assumptions needed for the Gronwall argument. The authors should either provide a complete proof of Eq. (22) or explicitly label it as a heuristic bound that motivates the empirical dissimilarity rather than a theorem.

minor comments (5)

[Supp. A.3.1] The text reads "We fill be using a Multi-layer Perceptron"; this should be "We will be using a multi-layer perceptron."
[Eq. (18)] The line "g(x) = f(x) target system" is confusing because g is introduced as the source system; the decomposition should use distinct labels or a clearer explanatory sentence.
[Supp. J.1.1] The paper should state explicitly whether the similarity scores in Figs. 3–4 are computed on the validation split used during training or on a fully independent test set; the text says "validation dataset" and this needs clarification for reproducibility.
[Supp. F.2.1] The initialization "MLP weight initialization sim N(0.02, 0.5)" is ambiguous; please specify whether 0.5 is the standard deviation or the variance.
[§3.1] The phrase "reasonable duration of time[0, Tmax]" is missing a space before the interval; a minor typo.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the DAA dissimilarity is defined as the fitted loss itself and is evaluated on held-out trajectories against a priori archetypes.

full rationale

The paper's derivation chain is self-contained. Equations (5) and (6) define the proposed dissimilarity as a two-part measure: a trajectory-matching loss under a coordinate diffeomorphism and a complexity penalty on that diffeomorphism. Equation (23) is explicitly presented as an empirical estimator of the same trajectory loss, fitted by a Neural ODE and evaluated on held-out data, so the reported similarity scores are normalized transforms of the fitted objective, not independent predictions masquerading as confirmations. The archetypes are defined a priori in Section 4 and the appendix, not inferred from the target data, and the numerical experiments compare the fitted losses across a fixed library. The only self-citations (refs. 19 and 20) provide background on continuous attractor fragility and supply the compass RNN target; they are not used to justify the method's validity, to restrict alternatives, or to import a uniqueness theorem. The Grönwall bound in Eq. (7) is an independent parameter-free inequality, and the decomposition in Eqs. (18)-(21) is definitional bookkeeping rather than a result derived from its own conclusion. The known non-invariance of Eqs. (5)-(6) under rotations and reflections is a correctness and robustness concern, not a circularity: the measure is what it is defined to be, and the paper does not claim to predict coordinate-frame-independent classifications without caveats.

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

The method's free parameters are all hand-chosen (Tmax, measure, norm, normalization, architecture), not fitted to data in a way that guarantees the reported classification; the standardization and similarity normalization are data-dependent and affect absolute values. The axioms are mostly standard (Gronwall, approximation theorems) plus domain assumptions about autonomous behavior and low-dimensional support; two ad hoc assumptions (existence of minimal conjugacy perturbation, first-order Ansatz) underpin the theoretical bounds but are not load-bearing for the empirical estimator.

free parameters (6)

Tmax = 2, 5, or 15 per experiment
The finite behavioral time horizon in Eq. (3) is chosen by hand for each target system (Supp. Tab. 4); the dissimilarity depends on it.
The sampling measure mu(x) = uniform over sampled trajectory points
The weighting over phase space is chosen as uniform over the observed trajectory points (Supp. A.4.1), affecting the expectations in Eqs. (4) and (5).
Complexity norm choice = Frobenius norm, p = 2
The deformation complexity in Eq. (6) uses a norm on JPhi - I; the choice of norm and p influences the complexity measurements.
Data standardization (mu, sigma) = computed from training data
Trajectories are centered and scaled per dimension before fitting (Supp. H.1), which changes the meaning of trajectory distance and diffeomorphism complexity.
Similarity score normalization = per-archetype max normalization
Dissimilarity scores are normalized by the maximum across targets for each archetype (Supp. J.1.1), turning absolute distances into relative rankings.
Neural ODE hidden sizes and epochs = [64] or [128], 200 or 1000 epochs
Architecture and training budget are hand-chosen and bound the expressiveness of the fitted diffeomorphism (Supp. G and I).

assumptions (6)

standard math Gronwall's inequality
Used in Eqs. (7) and (8) to bound trajectory deviation from a perturbed vector field.
standard math Universal approximation of homeomorphisms by Neural ODEs and i-ResNets
Cites ref. 60 to justify that the fitted diffeomorphism class is expressive enough.
domain assumption Autonomous dynamics over a finite interval [0, Tmax] with a known sampling distribution
The theory assumes no input and a known measure over initial conditions; the target RNN is treated as autonomous without stating how inputs were removed.
domain assumption The support of mu is contained in a low-dimensional, task-relevant region
The paper states that only utilized regions of phase space should be compared, and mu is taken over trajectory points.
ad hoc to paper Existence of a minimal perturbation achieving topological conjugacy on the support of mu
Used in Eqs. (9) through (11); no constructive proof is given, and the bound in Eq. (11) does not follow as stated.
ad hoc to paper The optimal diffeomorphism can be approximated by a first-order expansion Phi = x + epsilon*Phi_tilde + O(epsilon^2)
The bound in Eqs. (15) and (16) relies on this Ansatz, which is plausible but not justified for the fitted Neural ODE maps.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Dynamical Archetype Analysis: Autonomous Computation." pith.science (2026). https://pith.science/paper/K2HCT2PU

@misc{pith2026250705505,
  author       = {Pith},
  title        = {Pith review of: Dynamical Archetype Analysis: Autonomous Computation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K2HCT2PU}},
  note         = {Machine review of arXiv:2507.05505}
}

read the original abstract

The study of neural computation aims to understand the function of a neural system as an information processing machine. Neural systems are undoubtedly complex, necessitating principled and automated tools to abstract away details to organize and incrementally build intuition. We argue that systems with the same effective behavior should be abstracted by their ideal representative, i.e., archetype, defined by its asymptotic dynamical structure. We propose a library of archetypical computations and a new measure of dissimilarity that allows us to group systems based on their effective behavior by explicitly considering both deformations that break topological conjugacy as well as diffeomorphisms that preserve it. The proposed dissimilarity can be estimated from observed trajectories. Numerical experiments demonstrate our method's ability to overcome previously reported fragility of existing (dis)similarity measures for approximate continuous attractors and high-dimensional recurrent neural networks. Although our experiments focus on working memory systems, our theoretical approach naturally extends to general mechanistic interpretation of recurrent dynamics in both biological and artificial neural systems. We argue that abstract dynamical archetypes, rather than detailed dynamical systems, offer a more useful vocabulary for describing neural computation.

Figures

Figures reproduced from arXiv: 2507.05505 by the authors.

**Figure 1.** (A) Fixed points (red circles) characterize the asymptotic behavior and topological conjugacy (t.c.) in 1-D dynamical systems. Although the systems in the bottom are not t.c. to each other, yet they behave similarly in the short time scale. (B) Behavioral dissimilarity consists of two measures of distortion to map a source system to a target system. Their trade-off forms a boundary of minimum amount of distortion n… view at source ↗

**Figure 2.** Dissimilarity to an increasingly distorted ring attractor. ( [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Archetypes are assessed for each target system based on two criteria: their similarity to [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (18 more)

**Figure 4.** Figure 4: We compare our method to the state-of-the-art methods DSA [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: The Bistable archetype’s vector field and phase portrait. E.3 Limit cycle The limit cycle system is defined in polar coordinates as follows: r˙ = αr(r − 1), (42) ˙θ = v, (43) where α controls the rate of radial convergence to the unit circle, and v determines the angul…

**Figure 6.** Figure 6: The limit cycle archetype’s vector field and phase portrait. E.3.1 Analytical limit cycle Given the radial dynamics in Eq. 42 the solution is: r(t) = 1 1 + C′e−αt , where C ′ = 1 − r0 r0 . For the angular dynamics in Eq. 43 the solution is: ˙θ = v ⇒ θ(t) = vt + θ0 For …

**Figure 7.** Figure 7: The ring attractor archetype’s vector field and phase portrait. E.4.2 Bump-modulated perturbation on the ring modifying the angular dynamics from a static ˙θ = 0 to a bump-modulated perturbation on the ring. The dynamics are given by ˙θ = exp − (r − 1)2 2σ 2 · ψ(θ)…

**Figure 8.** Figure 8: Deterministic and noisy Van der Pol systems. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png]

**Figure 9.** Figure 9: Vector fields and trajectories of Sel’kov and Li [PITH_FULL_IMAGE:figures/full_fig_p027_9.png]

**Figure 10.** Figure 10: Deterministic and noisy Van der Pol systems. [PITH_FULL_IMAGE:figures/full_fig_p027_10.png]

**Figure 11.** Figure 11: The Two Bounded Line Attractors system’s vector field and used trajectories to train and evaluate. F.4 Compass RNN On each trial, the RNN receives a one-dimensional input sequence of angular velocities (positive for rightward, negative for leftward rotation), sampled …

**Figure 12.** Figure 12: Our method applied to the intermediate interpolation values for the deformed ring at [PITH_FULL_IMAGE:figures/full_fig_p033_12.png]

**Figure 13.** Figure 13: The Smooth Prototype Equivalences method applied to the deformed ring attractor as [PITH_FULL_IMAGE:figures/full_fig_p034_13.png]

**Figure 14.** Figure 14: Our method applied to the intermediate interpolation values for the perturbed ring at [PITH_FULL_IMAGE:figures/full_fig_p035_14.png]

**Figure 15.** Figure 15: The Smooth Prototype Equivalences (SPE30) method applied to the deformed ring attractor as described in Supp.Sec. F.2.2. The norm of the perturbation is shown above each subplot. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_15.png]

**Figure 16.** Figure 16: The ring attractor archetype fit to RNNs of different sizes ( [PITH_FULL_IMAGE:figures/full_fig_p037_16.png]

**Figure 17.** Figure 17: The ring attractor archetype fitted to the different target systems. 38 [PITH_FULL_IMAGE:figures/full_fig_p038_17.png]

**Figure 18.** Figure 18: The limit cycle archetype fitted to the different target systems. 39 [PITH_FULL_IMAGE:figures/full_fig_p039_18.png]

**Figure 19.** Figure 19: The single fixed point archetype fitted to the different target systems. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_19.png]

**Figure 20.** Figure 20: The bistable archetype fitted to the different target systems. 41 [PITH_FULL_IMAGE:figures/full_fig_p041_20.png]

**Figure 21.** Figure 21: The Bounded Line Attractor archetype fitted to the different target systems. 42 [PITH_FULL_IMAGE:figures/full_fig_p042_21.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Universal Approximation Theorems for Dynamical Systems with Infinite-Time Horizon Guarantees
math.DS 2026-02 conditional novelty 7.0 of 10

Neural ODEs can approximate Morse-Smale and continuous-attractor dynamical systems over infinite time in an ε-δ sense, provided limit-cycle periods are matched exactly.

Reference graph

Works this paper leans on

80 extracted references · 74 canonical work pages · cited by 1 Pith paper

[1]

J., Linderman, S

Nair, A., Karigo, T., Yang, B., Ganguli, S., Schnitzer, M. J., Linderman, S. W., Anderson, D. J. & Kennedy, A. An approximate line attractor in the hypothalamus encodes an aggressive state. Cell, 186(1):178–193. e15, 2023

work page 2023
[2]

C., Ye, J., Zoltowski, D

Pei, F. C., Ye, J., Zoltowski, D. M., Wu, A., Chowdhury, R. H., Sohn, H., O’Doherty, J. E., Shenoy, K. V ., Kaufman, M., Churchland, M. M. & others, . Neural Latents Benchmark ’21: Evaluating latent variable models of neural population activity. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021

work page 2021
[3]

& Park, I

Dowling, M., Zhao, Y . & Park, I. M. eXponential FAmily dynamical systems (XFADS): Large- scale nonlinear gaussian state-space modeling. In Advances in Neural Information Processing Systems (NeurIPS), December 2024

work page 2024
[4]

Machens, C. K. & Brody, C. D. Design of continuous attractor networks with monotonic tuning using a symmetry principle. Neural Computation, 20(2):452–485, 2008

work page 2008
[5]

Mante, V ., Sussillo, D., Shenoy, K. V . & Newsome, W. T. Context-dependent computation by recurrent dynamics in prefrontal cortex. Nature, 503(7474):78–84, 2013

work page 2013
[6]

N., Shenoy, K

Driscoll, L. N., Shenoy, K. & Sussillo, D. Flexible multitask computation in recurrent networks utilizes shared dynamical motifs. Nature Neuroscience, 27(7):1349–1363, 2024

work page 2024
[7]

D., Sussillo, D

Vyas, S., Golub, M. D., Sussillo, D. & Shenoy, K. V . Computation through neural population dynamics. Annual Review of Neuroscience, 43:249–275, 2020

work page 2020
[8]

D., Ostrow, M., Zoltowski, D

Versteeg, C., McCart, J. D., Ostrow, M., Zoltowski, D. M., Washington, C. B., Driscoll, L., Codol, O., Michaels, J. A., Linderman, S. W., Sussillo, D. & others, . Computation-through- Dynamics Benchmark: Simulated datasets and quality metrics for dynamical models of neural activity. bioRxiv, pages 2025–02, 2025

work page 2025

Show all 80 references

[9]

Beer, R. D. On the dynamics of small continuous-time recurrent neural networks. Adaptive Behavior, 3(4):469–509, 1995

1995
[10]

Beer, R. D. Parameter space structure of continuous-time recurrent neural networks. Neural Computation, 18(12):3009–3051, 2006

2006
[11]

Neural circuits as computational dynamical systems

Sussillo, D. Neural circuits as computational dynamical systems. Current opinion in neurobi- ology, 25:156–163, 2014

2014
[12]

Zhang, W., Wu, Y . N. & Wu, S. Translation-equivariant representation in recurrent networks with a continuous manifold of attractors. Advances in Neural Information Processing Systems, 35:15770–15783, 2022

2022
[13]

& Romani, S

Hoeller, J., Zhong, L., Pachitariu, M. & Romani, S. Bridging tuning and invariance with equivariant neuronal representations. bioRxiv, pages 2024–08, 2024

2024
[14]

& Miolane, N

Kurtkaya, B., Dinc, F., Yuksekgonul, M., Blanco-Pozo, M., Cirakman, E., Schnitzer, M., Yemez, Y ., Tanaka, H., Yuan, P. & Miolane, N. Dynamical phases of short-term memory mechanisms in RNNs. arXiv preprint arXiv:2502.17433, 2025

2025 arXiv
[15]

Strogatz, S. H. Nonlinear dynamics and chaos : with applications to physics, biology, chem- istry, and engineering. Studies in nonlinearity. Westview Press, 1 edition, January 2000. ISBN 9780738204536

2000
[16]

Arnol’d, V . I. The theory of bifurcations in the work of A.A. Andronov. In Dynamical Sys- tems V, Encyclopaedia of mathematical sciences, pages 224–232. Springer Berlin Heidelberg, Berlin, Heidelberg, 1994. ISBN 9783540653790,9783642578847

1994
[17]

A lemma in the theory of structural stability of differential equations

Hartman, P. A lemma in the theory of structural stability of differential equations. Proceedings of the American Mathematical Society, 11(4):610–620, 1960

1960
[18]

& Abbott, L

Dayan, P. & Abbott, L. F. Theoretical neuroscience: Computational and mathematical model- ing of neural systems. MIT press, 2005. 11

2005
[19]

& Park, I

S ´agodi, A., Mart´ın-S´anchez, G., Sok ´oł, P. & Park, I. M. Back to the continuous attractor. In Advances in Neural Information Processing Systems (NeurIPS), December 2024

2024
[20]

M., S ´agodi, ´A

Park, I. M., S ´agodi, ´A. & Sok´oł, P. A. Persistent learning signals and working memory without continuous attractors. August 2023, arXiv:2308.12585 [q-bio.NC]

2023 arXiv
[21]

& Bandettini, P

Kriegeskorte, N., Mur, M. & Bandettini, P. A. Representational similarity analysis-connecting the branches of systems neuroscience. Frontiers in systems neuroscience, 2:249, 2008

2008
[22]

& Sohl-Dickstein, J

Raghu, M., Gilmer, J., Yosinski, J. & Sohl-Dickstein, J. SVCCA: Singular vector canoni- cal correlation analysis for deep learning dynamics and interpretability. Advances in Neural Information Processing Systems, 30, 2017

2017
[23]

H., Kunz, E., Kornblith, S

Williams, A. H., Kunz, E., Kornblith, S. & Linderman, S. Generalized shape metrics on neural representations. Advances in Neural Information Processing Systems, 34:4738–4750, 2021

2021
[24]

R., Zhou, J., Nassar, J., Berman, J., Olieslagers, J

Duong, L. R., Zhou, J., Nassar, J., Berman, J., Olieslagers, J. & Williams, A. H. Rep- resentational dissimilarity metric spaces for stochastic neural networks. arXiv preprint arXiv:2211.11665, 2022

2022 arXiv
[25]

R., Harvey, S

Barbosa, J., Nejatbakhsh, A., Duong, L. R., Harvey, S. E., Brincat, S. L., Siegel, M., Miller, E. K. & Williams, A. H. Quantifying differences in neural population activity with shape metrics. bioRxiv, pages 2025–01, 2025

2025
[26]

Williams, A. H. Equivalence between representational similarity analysis, centered kernel alignment, and canonical correlations analysis. bioRxiv, pages 2024–10, 2024

2024
[27]

& Fiete, I

Ostrow, M., Eisen, A., Kozachkov, L. & Fiete, I. Beyond geometry: Comparing the temporal structure of computation in neural circuits with dynamical similarity analysis. Advances in Neural Information Processing Systems, 36, 2024

2024
[28]

& Nitzan, M

Moriel, N., Ricci, M. & Nitzan, M. Let’s do the time-warp-attend: Learning topological invariants of dynamical systems. In The Twelfth International Conference on Learning Repre- sentations, 2024

2024
[29]

Chen, R., Vedovati, G., Braver, T. S. & Ching, S. Dform: Diffeomorphic vector field alignment for assessing dynamics across learned models. CoRR, abs/2402.09735, 2024

2024 arXiv
[30]

& Nitzan, M

Friedman, R., Moriel, N., Ricci, M., Pelc, G., Weiss, Y . & Nitzan, M. Characterizing nonlinear dynamics via smooth prototype equivalences. arXiv preprint arXiv:2503.10336, 2025

2025
[31]

& Ortega, J.-P

Grigoryeva, L. & Ortega, J.-P. Dimension reduction in recurrent networks by canonicalization. arXiv preprint arXiv:2007.12141, 2020

2007 arXiv
[32]

& Ortega, J.-P

Gonon, L., Grigoryeva, L. & Ortega, J.-P. Approximation bounds for random neural networks and reservoir systems. The Annals of Applied Probability, 33(1):28–69, 2023

2023
[33]

& Thurm, M

Durstewitz, D., Koppe, G. & Thurm, M. I. Reconstructing computational system dynamics from neural data with recurrent neural networks. Nature Reviews Neuroscience, 24(11):693– 710, 2023

2023
[34]

J., Monfared, Z

Brenner, M., Hemmer, C. J., Monfared, Z. & Durstewitz, D. Almost-linear RNNs yield highly interpretable symbolic codes in dynamical systems reconstruction. arXiv preprint arXiv:2410.14240, 2024

2024 arXiv
[35]

& Sahani, M

Duncker, L., Bohner, G., Boussard, J. & Sahani, M. Learning interpretable continuous-time models of latent stochastic dynamical systems. In International conference on machine learn- ing, pages 1726–1734. PMLR, 2019

2019
[36]

& Haller, G

Cenedese, M., Ax ˚as, J., B ¨auerlein, B., Avila, K. & Haller, G. Data-driven modeling and prediction of non-linearizable dynamics via spectral submanifolds. Nature communications, 13(1):872, 2022

2022
[37]

& Nitzan, M

Ricci, M., Moriel, N., Piran, Z. & Nitzan, M. Phase2vec: Dynamical systems embedding with a physics-informed convolutional network. arXiv preprint arXiv:2212.03857, 2022. 12

2022 arXiv
[38]

L., Arnaudon, A., Barahona, M

Gosztolai, A., Peach, R. L., Arnaudon, A., Barahona, M. & Vandergheynst, P. MARBLE: Interpretable representations of neural population dynamics using geometric deep learning. Nature Methods, 2025

2025
[39]

The Gronwall inequality

Howard, R. The Gronwall inequality. arXiv [math.FA], March 2025, arXiv:2503.23639 [math.FA]

2025 arXiv
[40]

& Barak, O

Sussillo, D. & Barak, O. Opening the black box: Low-dimensional dynamics in high- dimensional recurrent neural networks. Neural Computation, 25(3):626–649, 2013

2013
[41]

Katz, G. E. & Reggia, J. A. Using directional fibers to locate fixed points of recurrent neural networks. IEEE transactions on neural networks and learning systems , 29(8):3636–3646, 2017

2017
[42]

Golub, M. D. & Sussillo, D. FixedPointFinder: A Tensorflow toolbox for identifying and characterizing fixed points in recurrent neural networks. Journal of Open Source Software, 3 (31):1003, 2018

2018
[43]

G., McClements, W., Ruiz, A

Townley, S., Ilchmann, A., Weiß, M. G., McClements, W., Ruiz, A. C., Owens, D. H. & Pratzel-Wolters, D. Existence and learning of oscillations in recurrent neural networks. IEEE Transactions on neural networks, 11(1):205–214, 2000

2000
[44]

E., Pei, F., Gloeckler, M

Pals, M., Sa ˘gtekin, A. E., Pei, F., Gloeckler, M. & Macke, J. H. Inferring stochastic low-rank recurrent neural networks from neural data. arXiv preprint arXiv:2406.16749, 2024

2024 arXiv
[45]

The Brain Abstracted: Simplification in the history and philosophy of neuro- science

Chirimuuta, M. The Brain Abstracted: Simplification in the history and philosophy of neuro- science. MIT Press, 2024

2024
[46]

F., DePasquale, B., Pillow, J

Panichello, M. F., DePasquale, B., Pillow, J. W. & Buschman, T. J. Error-correcting dynamics in visual working memory. Nature communications, 10(1):3366, 2019

2019
[47]

M., Ardalan, A., Markov, N

Tafazoli, S., Bouchacourt, F. M., Ardalan, A., Markov, N. T., Uchimura, M., Mattar, M. G., Daw, N. D. & Buschman, T. J. Building compositional tasks with shared neural subspaces. bioRxiv, 2024

2024
[48]

& Engel, T

Langdon, C., Genkin, M. & Engel, T. A. A unifying perspective on neural manifolds and circuits for cognition. Nature Reviews Neuroscience, pages 1–15, 2023

2023
[49]

& Krishnamurthy, K

Can, T. & Krishnamurthy, K. Emergence of memory manifolds. arXiv preprint arXiv:2109.03879, 2021

2021 arXiv
[50]

J., Ardalan, A., Tsodyks, M

Cueva, C. J., Ardalan, A., Tsodyks, M. & Qian, N. Recurrent neural network models for work- ing memory of continuous variables: Activity manifolds, connectivity patterns, and dynamic codes. arXiv preprint arXiv:2111.01275, 2021

2021 arXiv
[51]

Emergence of universal computations through neural manifold dynamics

Gort, J. Emergence of universal computations through neural manifold dynamics. Neural Computation, 36(2):227–270, 2024

2024
[52]

& Narayanan, R

Mishra, P. & Narayanan, R. Stable continual learning through structured multiscale plasticity manifolds. Current opinion in neurobiology, 70:51–63, 2021

2021
[53]

& Fiete, I

Chaudhuri, R., Gerc ¸ek, B., Pandey, B., Peyrache, A. & Fiete, I. The intrinsic attractor manifold and population dynamics of a canonical cognitive circuit across waking and sleep. Nature neuroscience, 22(9):1512–1520, 2019

2019
[54]

& Ching, S

Ghazizadeh, E. & Ching, S. Slow manifolds within network dynamics encode working mem- ory efficiently and robustly. PLoS computational biology, 17(9):e1009366, 2021

2021
[55]

& Sahani, M

Duncker, L. & Sahani, M. Dynamics on the manifold: Identifying computational dynamical activity from neural population recordings. Current opinion in neurobiology , 70:163–170, 2021

2021
[56]

& Gerstner, W

Pezon, L., Schmutz, V . & Gerstner, W. Linking neural man- ifolds to circuit structure in recurrent networks. bioRxiv, 2024, https://www.biorxiv.org/content/early/2024/06/10/2024.02.28.582565.full.pdf. 13

2024
[57]

C., Miller, L

Fortunato, C., Bennasar-V ´azquez, J., Park, J., Chang, J. C., Miller, L. E., Dudman, J. T., Perich, M. G. & Gallego, J. A. Nonlinear manifolds underlie neural population activity during behaviour. bioRxiv, pages 2023–07, 2024

2023
[58]

T., Rubanova, Y ., Bettencourt, J

Chen, R. T., Rubanova, Y ., Bettencourt, J. & Duvenaud, D. K. Neural ordinary differential equations. Advances in Neural Information Processing Systems, 31, 2018

2018
[59]

Popper, K. R. The open universe: An argument for indeterminism , volume 2. Psychology Press, 1988

1988
[60]

& Arodz, T

Zhang, H., Gao, X., Unterman, J. & Arodz, T. Approximation capabilities of Neural ODEs and Invertible Residual Networks. In III, H. D. & Singh, A., editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Rese...

2020
[61]

Cohen, M. A. & Grossberg, S. Absolute stability of global pattern formation and parallel memory storage by competitive neural networks. IEEE transactions on systems, man, and cybernetics, (5):815–826, 1983

1983
[62]

D., Luo, T

Kim, T. D., Luo, T. Z., Pillow, J. W. & Brody, C. D. Inferring latent dynamics underlying neural population activity via neural differential equations. In International Conference on Machine Learning, pages 5551–5561. PMLR, 2021

2021
[63]

& Calhoun, V

Geenjaar, E., Kim, D., Ohib, R., Duda, M., Kashyap, A., Plis, S. & Calhoun, V . Learn- ing low-dimensional dynamics from whole-brain data improves task capture. arXiv preprint arXiv:2305.14369, 2023

2023 arXiv
[64]

R., Versteeg, C

Sedler, A. R., Versteeg, C. & Pandarinath, C. Expressive architectures enhance interpretability of dynamics-based neural population models. Neurons, behavior, data analysis, and theory , 2023

2023
[65]

& Gerven, M.van

ElGazzar, A. & Gerven, M.van . Universal differential equations as a common modeling language for neuroscience. arXiv preprint arXiv:2403.14510, 2024

2024 arXiv
[66]

Rubanova, Y ., Chen, R. T. & Duvenaud, D. K. Latent ordinary differential equations for irregularly-sampled time series. Advances in Neural Information Processing Systems , 32, 2019

2019
[67]

Coelho, C., Costa, M. F. P. & Ferr´as, L. L. Enhancing continuous time series modelling with a latent ODE-LSTM approach. Applied Mathematics and Computation, 475:128727, 2024

2024
[68]

& Koseska, A

Koch, D., Nandan, A., Ramesan, G. & Koseska, A. Biological computations: Limitations of attractor-based formalisms and the need for transients. Biochemical and Biophysical Research Communications, 720:150069, 2024

2024
[69]

Skufca, J. D. & Bollt, E. M. Relaxing conjugacy to fit modeling in dynamical systems.Physical Review E—Statistical, Nonlinear, and Soft Matter Physics, 76(2):026220, 2007

2007
[70]

Skufca, J. D. & Bollt, E. M. A concept of homeomorphic defect for defining mostly conjugate dynamical systems. Chaos: An Interdisciplinary Journal of Nonlinear Science, 18(1), 2008

2008
[71]

Bollt, E. M. & Skufca, J. D. On comparing dynamical systems by defective conjugacy: A symbolic dynamics interpretation of commuter functions. Physica D: Nonlinear Phenomena, 239(10):579–590, 2010

2010
[72]

& Park, I

Zhao, Y . & Park, I. M. Interpretable nonlinear dynamic modeling of neural trajectories. In Advances in Neural Information Processing Systems (NIPS), 2016

2016
[73]

W., Bugallo, M

Nassar, J., Linderman, S. W., Bugallo, M. & Park, I. M. Tree-structured recurrent switching linear dynamical systems for multi-scale modeling. In International Conference on Learning Representations (ICLR), November 2019

2019
[74]

On ’relaxation-oscillations’

Pol, B.van der . On ’relaxation-oscillations’. The London, Edinburgh, and Dublin Philosophi- cal Magazine and Journal of Science, 2(11):978–992, 1926. 14

1926
[75]

Sel’kov, E. E. Self-oscillations in glycolysis 1. A simple kinetic model. European Journal of Biochemistry, 4(1):79–86, 1968

1968
[76]

& Asama, H

Massaroli, S., Poli, M., Park, J., Yamashita, A. & Asama, H. Dissecting Neural ODEs. Ad- vances in Neural Information Processing Systems, 33:3952–3963, 2020

2020
[77]

Chen, R. T. Q. torchdiffeq, 2018. Available from: https://github.com/rtqichen/ torchdiffeq

2018
[78]

closeness

Kingma, D. P. & Ba, J. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 15 A Topology A.1 Approximating homeomorphisms Any homeomorphism on a p-dimensional Euclidean space can be approximated by a Neural ODE or an i-ResNet operating on a 2p-dim...

2014 arXiv
[79]

Inside the Bounded Region When the state x is within the hypercube [−B, B]D, the flow is zero: dx dt = 0, for x ∈ [−B, B]D
[80]

The projection proj (x) can be written component-wise as: proj(xi) = clip(xi, −B, B) for each i = 1, 2,

Outside the Bounded Region When the state x is outside the hypercube [−B, B]D, the flow is directed towards the nearest bound- ary, with a scaling factor α controlling the flow magnitude: dx dt = −α(x − proj(x)), for x /∈ [−B, B]D, where proj(x) denotes the projection of x ont...

1968

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

J., Linderman, S

Nair, A., Karigo, T., Yang, B., Ganguli, S., Schnitzer, M. J., Linderman, S. W., Anderson, D. J. & Kennedy, A. An approximate line attractor in the hypothalamus encodes an aggressive state. Cell, 186(1):178–193. e15, 2023

work page 2023

[2] [2]

C., Ye, J., Zoltowski, D

Pei, F. C., Ye, J., Zoltowski, D. M., Wu, A., Chowdhury, R. H., Sohn, H., O’Doherty, J. E., Shenoy, K. V ., Kaufman, M., Churchland, M. M. & others, . Neural Latents Benchmark ’21: Evaluating latent variable models of neural population activity. In Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track (Round 2), 2021

work page 2021

[3] [3]

& Park, I

Dowling, M., Zhao, Y . & Park, I. M. eXponential FAmily dynamical systems (XFADS): Large- scale nonlinear gaussian state-space modeling. In Advances in Neural Information Processing Systems (NeurIPS), December 2024

work page 2024

[4] [4]

Machens, C. K. & Brody, C. D. Design of continuous attractor networks with monotonic tuning using a symmetry principle. Neural Computation, 20(2):452–485, 2008

work page 2008

[5] [5]

Mante, V ., Sussillo, D., Shenoy, K. V . & Newsome, W. T. Context-dependent computation by recurrent dynamics in prefrontal cortex. Nature, 503(7474):78–84, 2013

work page 2013

[6] [6]

N., Shenoy, K

Driscoll, L. N., Shenoy, K. & Sussillo, D. Flexible multitask computation in recurrent networks utilizes shared dynamical motifs. Nature Neuroscience, 27(7):1349–1363, 2024

work page 2024

[7] [7]

D., Sussillo, D

Vyas, S., Golub, M. D., Sussillo, D. & Shenoy, K. V . Computation through neural population dynamics. Annual Review of Neuroscience, 43:249–275, 2020

work page 2020

[8] [8]

D., Ostrow, M., Zoltowski, D

Versteeg, C., McCart, J. D., Ostrow, M., Zoltowski, D. M., Washington, C. B., Driscoll, L., Codol, O., Michaels, J. A., Linderman, S. W., Sussillo, D. & others, . Computation-through- Dynamics Benchmark: Simulated datasets and quality metrics for dynamical models of neural activity. bioRxiv, pages 2025–02, 2025

work page 2025