Pith. sign in

REVIEW 3 major objections 5 minor 7 references

The paper claims that a low-rank recurrent neural network, reframed as a variational autoencoder with a partial-correlation penalty, can disentangle neural latent dynamics into independent groups and decompose the connectivity matrix into i

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

DisRNN, a VAE-based low-rank RNN with a group-wise independence penalty, learns disentangled latent trajectories and interpretable sub-connectivity from neural population recordings.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Useful VAE-based method for group-wise disentanglement of low-rank RNN latents, but the paper overclaims process-level independence from a marginal penalty. the 3 major comments →

arxiv 2511.13899 v2 pith:E4674UVG submitted 2025-11-17 q-bio.NC cs.CEcs.LG

A Factorized Low-Rank RNN Framework for Uncovering Independent Neural Latent Dynamics and Connectivity

Chengrui Li , Yunmiao Wang , Yule Wang , Weihan Li , Dieter Jaeger , Anqi Wu This is my paper
classification q-bio.NC cs.CEcs.LG
keywords low-rank recurrent neural networklatent disentanglementpartial correlationvariational autoencoderneural population dynamicsconnectivity decompositionmotor cortexvoltage imaging
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 tries to show that neural population recordings can be explained by a low-rank recurrent network whose latent dynamics split into a few independent groups, each with its own sub-connectivity. The authors reformulate the low-rank RNN as a variational autoencoder and add a partial-correlation penalty that pushes the latent groups toward independence without forcing individual dimensions to be independent. They claim this yields better disentanglement of neural trajectories and interpretable connectivity decompositions on synthetic, monkey motor cortex, and mouse voltage-imaging data. The importance is that it offers a way to assign distinct computational roles to separate neural sub-circuits.

Core claim

The central claim is that group-wise independence, rather than full component-wise independence, is the right structural assumption for latent dynamics in low-rank RNNs. By expressing the low-rank RNN in a variational autoencoder framework, the paper introduces a partial-correlation penalty on the aggregated posterior that encourages independence between groups of latent dimensions while allowing flexible within-group entanglement. This allows the rank-K connectivity W = AB to be decomposed into sub-connectivities W_g = A_{:,g}B_{g,:}, each associated with a distinct latent group, which the paper interprets as neural sub-circuits. Experiments on synthetic data, macaque M1 recordings, and mou

What carries the argument

The central mechanism is the partial-correlation penalty added to the evidence lower bound objective of a variational autoencoder reformulation of the low-rank RNN. The penalty, scaled by a hyperparameter beta, penalizes the KL divergence between the aggregated posterior q(z) and the product of group marginals prod_g q(z_g). This encourages group-wise independence without forcing individual dimensions within a group to be independent. The second key piece is the decomposition of the learned low-rank connectivity into group-specific sub-connectivities W_g = A_{:,g}B_{g,:}, which gives the interpretation of separate neural sub-circuits.

Load-bearing premise

The load-bearing premise is that the true neural population actually has separable latent groups—that minimizing the partial-correlation penalty recovers the true independent groups rather than an arbitrary near-independent transformation of them, and that the model's group structure matches a real biological decomposition.

What would settle it

A concrete check would be to generate synthetic data from a low-rank RNN with known but entangled latent groups (i.e., no true independent groups) and see if the proposed model still forces a group-wise decomposition with high confidence; if it does, the penalty can create artifacts. Alternatively, on real data, compare the recovered sub-connectivities to known synaptic or anatomical connectivity via independent measurements; a mismatch would challenge the sub-circuit interpretation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Neural circuits can be decomposed into independent sub-circuits that drive separate behavioral or sensory variables, enabling circuit-level hypotheses from population recordings.
  • Low-rank RNN models can be made more interpretable without sacrificing reconstruction accuracy, since group-wise independence allows higher within-group rank.
  • The approach provides a way to identify which latent groups correspond to task-relevant variables (e.g., x/y movement coordinates) even without explicit labels.
  • The framework extends to nonlinear encoders/decoders and general nonlinear dynamical systems, suggesting broad applicability beyond linear connectivity.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If group-wise independence holds, the partial-correlation penalty might be seen as a form of identifiable factor analysis for dynamical systems, but identifiability beyond linear-Gaussian cases remains an open question the paper does not settle.
  • The paper's evidence for biological plausibility (preferred cardinal axes in M1) suggests a testable prediction: the recovered sub-connectivities should align with known anatomical or functional connectivity, which could be validated with optogenetic or pharmacological perturbations.
  • The method might be extended to incorporate task inputs or stimulus labels, which the paper notes as a limitation; such extensions could reveal input-driven changes in sub-circuit connectivity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes DisRNN, a VAE-based reformulation of low-rank RNNs in which a penalty on the KL divergence between the aggregated posterior and the product of its group marginals encourages group-wise independence among latent dimensions. The resulting model yields a decomposition W=AB=Σ_g A_{:,g}B_{g,:}, whose terms are interpreted as sub-connectivities associated with independent latent groups. The authors evaluate the method on a synthetic Lorenz/Thomas dataset, macaque M1 spiking data, and mouse dorsal cortex voltage imaging, comparing against lrRNN, LINT, SMC, and variants. They report improved latent alignment, better disentanglement, and more interpretable connectivity decompositions. The framework is coherent as a variational inference scheme, and the paper includes useful ablations and a rotated-coordinate check on M1. However, the central claim that DisRNN uncovers 'independent latent dynamics' is not fully supported by the training objective or the evaluations, and the synthetic data generation is under-specified.

Significance. If the central claim holds, the paper would make a useful contribution by extending low-rank RNN decompositions from orthogonal (SVD-based) to independent groups, enabling functional assignment of neural sub-circuits. The VAE reformulation is clean and the idea of partial disentanglement (rather than full dimension-wise independence) is well motivated. The synthetic ground-truth alignment and the rotated-coordinate analysis are constructive, and the β ablation is helpful. However, the manuscript currently does not establish the process-level independence it advertises, and the synthetic experiment—the main external validation—needs clarification. The paper also frankly acknowledges its own limitations (pre-specified group counts, linear connectivity, no external inputs), which is to its credit. Overall, the framework is promising but the evidence does not yet support the strongest claims.

major comments (3)
  1. [Sec. 3.2, Eq. (10), Eq. (9), Abstract] The independence penalty in Eq. (10) is KL(q(z)||Π_g q(z_g)), where q(z) is the time-averaged aggregated posterior. This is a marginal, zero-lag independence criterion: it can vanish even when z_g(t) depends on z_{g'}(t−l) for l>0. Eq. (9) also defines disentanglement through the marginal p(z), not the full process p(z_{1:T}). Therefore the Abstract's claim that DisRNN recovers 'independent latent dynamics' that 'evolve separately' is not supported by the training objective. The evaluations in Sec. 4.1 (PC, R² alignment) and Sec. 4.2 (rotated-coordinate alignment) are all zero-lag and cannot detect temporal cross-talk. Please add a lagged-dependence metric (e.g., lagged cross-correlation or transfer entropy) or modify the penalty/architecture to enforce process-level independence (e.g., a block-diagonal latent transition), and adjust the claims accordingly.
  2. [Sec. 4.1, Dataset] The synthetic generation procedure is underspecified and appears contradictory. The text first states 'To generate the latent, we simulate Lorenz and Thomas’ cyclically symmetric dynamics,' then says 'Since z(t) follows the generative process in Eq. (7), we can get the parameters including A, B, and hence generate the observed data x(t) using the fitted A and B.' A Lorenz/Thomas trajectory does not, in general, satisfy Eq. (7) for arbitrary A,B; conversely, if z(t) is generated from Eq. (7), it is unclear how the Lorenz/Thomas dynamics are used. This ambiguity undermines the interpretation of the ground-truth ‘recovery’ results. Please provide the exact generative scheme, including how A and B are chosen and how z(t) is made consistent with Eq. (7), or re-run the synthetic experiment with a properly specified generative process.
  3. [Sec. 3.2, Eq. (11)–(12)] The decomposition W=AB=Σ_g A_{:,g}B_{g,:} is an algebraic identity for any A and B; it does not by itself imply that each W_g is a 'sub-circuit' driving an independent source. That interpretation requires process-level independence of the latent groups, which is not guaranteed by the marginal PC penalty in Eq. (10). Consequently, the real-data connectivity interpretations in Secs. 4.2 and 4.3 (e.g., horizontal vs. vertical sub-networks, somatosensory sub-circuits) are post hoc descriptive decompositions rather than established independent sub-circuits. The paper should either enforce process-level independence (e.g., using bdRNN's block-diagonal J on the linear case) or explicitly frame W_g as an algebraic decomposition whose functional independence remains to be verified.
minor comments (5)
  1. [Eq. (10)] The quantity KL(q(z)||Π_g q(z_g)) is the total correlation (multi-information) between groups, not a 'partial correlation' as named. Please use terminology that matches the quantity or justify the name.
  2. [Sec. 4.2] The text refers to 'our code for details' but no code or data link is provided. Given the complexity of the method and the need for reproducibility, please include a public release or a clear availability statement.
  3. [Sec. 4.2, K=2] The paper concedes that the K=2 trajectories 'remain visually poor' despite an alignment R² of about 0.65. Please address the visual quality (e.g., different scaling, plotting choices) and ensure the quantitative claim is not overstated in the abstract or main text.
  4. [Appendix A.2.1, Eq. (19)] The combinatorial count in Eq. (19) appears off by a factor of G!. For K=GH components partitioned into G unlabeled groups of size H, the standard formula is K!/(G!(H!)^G), not the expression given (which evaluates to K! G!/(H!)^G). Please verify the calculation.
  5. [Sec. 4.3] Interpretations such as 'potentially indicating the formation of episodic memory of receiving the airpuff' are highly speculative. Please phrase these as hypotheses and avoid overstating the evidence from a single voltage-imaging dataset.

Circularity Check

1 steps flagged

PC penalty is both the training objective and the primary disentanglement metric, but external ground-truth and behavioral alignments keep the central claim partially independent.

specific steps
  1. fitted input called prediction [Sec. 3.2, 'Inference via partial correlation' (Eq. 10); Sec. 4.1, 'Latent evaluations']
    "Following Li et al. (2025), we achieve group-wise independence by optimizing the target function L= 1/T XT t=1 ELBO(x(t)) −β·KL(q(z) || QG g=1 q(zg)). ... To evaluate the estimated latent unsupervisedly, we compute the PC of the estimated latent on the test set to check whether different methods uncover desired group structures. Fig. 2(a) shows that DisRNN achieves the lowest PC, indicating that it successfully recovers partially disentangled latent dynamics."

    The PC term in Eq. (10) is exactly the KL divergence between the aggregated posterior and its group-factorized product. The Sec. 4.1 evaluation computes the same PC on test latents and interprets a lower value as recovering the desired group structure. Since training explicitly minimizes this KL, the metric is the optimization target; hence 'DisRNN achieves the lowest PC' is largely a statement about fitting the regularizer, not independent confirmation of true independent groups. Synthetic ground-truth R2/connectivity and M1 behavioral alignment provide external anchors, so this is partial rather than complete circularity.

full rationale

Most of the derivation is algebraic and self-contained: the VAE reformulation (Eqs. 4-8), the block-diagonal bdRNN construction (Appendix A.1), and the sub-connectivity decomposition W=AB=Σ W_g (Eq. 12) are identities or definitions rather than empirical predictions. The central caveat is that the PC term in Eq. (10) is the same KL(q(z)||∏ q(z_g)) used in Sec. 4.1 as the main disentanglement metric, so 'lowest PC' is in part a check that the regularizer was optimized. This is a genuine but partial circularity. It is mitigated by external anchors: synthetic latent R^2 alignment to ground-truth groups and connectivity correlation, and M1 alignment to hand kinematics with a rotated-coordinate control, all of which do not reduce to the PC objective. The Li et al. (2025) reference is a self-citation, but the PC penalty is defined in this paper, so the citation is provenance rather than a load-bearing self-citation chain. The process-independence concern (marginal vs. temporal independence) is a correctness/validity risk rather than a circularity, since the paper does not claim to derive process-level independence from the PC objective. The paper's own limitations section notes that connectivity interpretability is restricted to linear encoder/decoder settings, another caveat that does not itself constitute circularity. Overall, the central claim retains independent content, so score 4 is appropriate.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The central claim rests on a small set of modeling choices: a linear low-rank RNN observation model, a pre-specified group structure, and a PC penalty whose strength is set by validation. The paper does not provide formal identifiability conditions for recovering true latent groups, and several hyperparameters (K, G, H, beta, psi) are chosen by the user rather than learned. The contributions are therefore conditional on those choices being correct for the data.

free parameters (4)
  • beta (partial-correlation penalty weight) = 20 (synthetic validation; used across datasets)
    Eq. (10); cross-validated in Appendix A.2.1 Fig. 6. Controls the strength of group independence and is a hand-set hyperparameter.
  • Group count G and group ranks H_g = (2,3) synthetic; (2,1)/(2,2) M1; (6,1)/(6,2) and grid on mouse
    Pre-specified in Sec. 3.2 and Sec. 4; the paper itself calls group pre-specification a limitation. The appendix shows sensitivity to (2,4) and (3,4).
  • Latent dimensionality K = 2, 4, 6, 12 in experiments
    Chosen per dataset; not determined by the model.
  • History convolution kernel psi_l
    Entered in Eq. (4)/(6) with sum-to-one constraint, but the learning rule or fixed values are not described in the text.
axioms (5)
  • domain assumption The observed neural population activity is generated by a low-rank RNN with a linear encoder/decoder structure (Eq. (6)).
    The entire framework and the connectivity interpretation W=AB depend on this linear observation model.
  • domain assumption Group-wise independence is the correct factorization of the latent dynamics.
    The paper assumes true latent dynamics split into independent groups, each internally correlated; used in Eq. (9) and all experiments.
  • standard math ELBO and the Gaussian KDE approximation of the aggregated posterior yield a valid partial-correlation penalty.
    Eq. (3), Eq. (10), and Makhzani et al. (2015) are invoked; this is standard variational inference, though the KDE bandwidth and finite-sample behavior are not analyzed.
  • standard math Jordan normal form argument for bdRNN (Appendix A.1).
    The block-diagonal latent recurrence is justified by the Jordan decomposition of a linear recurrence, which is mathematically standard.
  • ad hoc to paper The penalty strength beta and the Gaussian likelihood are reasonable for the datasets used.
    beta=20 is selected by validation on synthetic data and transferred to real data; the Gaussian observation noise in Eq. (5) is not independently checked on spike or voltage data.
invented entities (2)
  • Group-wise independent latent subspaces (latent groups) no independent evidence
    purpose: Model constructs that assign distinct computational roles to independent groups of latent dimensions.
    The model is fit to make groups independent; on synthetic data groups are known by construction, and on real data only behavioral correlation is shown, with no external perturbation or formal identifiability.
  • Sub-connectivity matrices W_g = A_{:,g}B_{g,:} as neural sub-circuits no independent evidence
    purpose: Interpret each latent group as a distinct low-rank neural sub-circuit in the connectivity matrix.
    These are derived from fitted A and B; the mouse interpretations (e.g., S2 to M2-m) are post hoc and not validated by causal experiments.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of A Factorized Low-Rank RNN Framework for Uncovering Independent Neural Latent Dynamics and Connectivity." pith.science (2026). https://pith.science/paper/E4674UVG

@misc{pith2026251113899,
  author       = {Pith},
  title        = {Pith review of: A Factorized Low-Rank RNN Framework for Uncovering Independent Neural Latent Dynamics and Connectivity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E4674UVG}},
  note         = {Machine review of arXiv:2511.13899}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Low-rank recurrent neural networks (lrRNNs) are a class of models that uncover low-dimensional latent dynamics underlying neural population activity. Although their functional connectivity is low-rank, it lacks independence interpretations, making it difficult to assign distinct computational roles to different latent dimensions. To address this, we propose the Factored Recurrent Neural Network (FacRNN), a generative lrRNN framework that assumes group-wise independence among latent dynamics while allowing flexible within-group entanglement. These independent latent groups allow latent dynamics to evolve separately, but are internally rich for complex computation. We reformulate the lrRNN under a variational autoencoder (VAE) framework, enabling us to introduce a partial correlation penalty that encourages independence between groups of latent dimensions. Experiments on synthetic, monkey M1, and mouse voltage imaging data show that FacRNN consistently improves the disentanglement and interpretability of learned neural latent trajectories in low-dimensional space and low-rank connectivity over baseline lrRNNs that do not encourage group-wise independence.

Figures

Figures reproduced from arXiv: 2511.13899 by Anqi Wu, Chengrui Li, Dieter Jaeger, Weihan Li, Yule Wang, Yunmiao Wang.

Figure 1
Figure 1. Figure 1: (a): Schematic of the DisRNN, showing independent latent groups and the corresponding low-rank connectivities. (b): Group-wise independence: (z1, z2) ⊥ (z3, z4), while within-groups are highly entangled, and components from different groups are marginally independent. Inference. We introduce two inference approaches to achieve partial disentanglement: one tailored for linear dynamics (Eq. (8)); while the o… view at source ↗
Figure 2
Figure 2. Figure 2: (a): The PC and R2 of the estimated latent, and the connectivity correlation. The star￾bars indicate the pairwise t-test significance levels. Arrows indicate the higher or lower the better. (b): Group 2 latent trajectories with the true dynamics in 3D plots; and the 1D trajectories from different methods on selected latent components. All latent dimensions are plotted in [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 3
Figure 3. Figure 3: (a): Recovered latent trajectories from different methods v.s. the ground-truth trajectories, and their alignment R2 scores. (b): The corresponding low-rank connectivity for K = 2 DisRNN and K = 4 DisRNN. For DisRNN, for example, the rank-4 connectivity can be decomposed into two rank-2 sub-connectivities, one responsible for the dynamics of the x-movement, and the other for the y-movement. (c): The R2 sco… view at source ↗
Figure 4
Figure 4. Figure 4: (a): Experiment setup. (b): Reconstruction performances w.r.t. different numbers of components for DisRNN or different numbers of groups for DisRNN. (c): Cortical map with region abbreviations. (d): 1 rank-12 group, a no disentanglement configuration, i.e., a standard low-rank RNN. Left are brain maps A:,g and the corresponding time series z (t) g . Right is the corresponding rank-12 connectivity correspon… view at source ↗
Figure 5
Figure 5. Figure 5: The synthetic latent consists of two independent groups. [PITH_FULL_IMAGE:figures/full_fig_p014_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Metrics w.r.t. the PC penalty β in DisRNN on the synthetic dataset. To verify that DisRNN is flexible in setting the number of groups G and group rank H, we run DisRNN with (G, H) = (2, 4) and (G, H) = (3, 4) against the true (G, H) = (2, 3), and evaluate the latent alignment outcome. The latent R2 for them are 0.84 ± 0.14 and 0.78 ± 0.09, respec￾tively. All of them are better than other baseline methods, … view at source ↗
Figure 7
Figure 7. Figure 7: The latent dynamics alignment process. We try to align the estimated latent groups with [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The estimated latent components by different methods on every latent compo [PITH_FULL_IMAGE:figures/full_fig_p016_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Left: Complexity demonstration matrix. Right: The latent R2 of all possible permutations for methods without the assumption of group structure under different random seeds. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The learned connectivity matrices from all methods in Fig. 3(a), except for [PITH_FULL_IMAGE:figures/full_fig_p017_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

7 extracted references · 6 linked inside Pith

  1. [5]

    Adversarial autoencoders.arXiv preprint arXiv:1511.05644,

    Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, Ian Goodfellow, and Brendan Frey. Adversarial autoencoders.arXiv preprint arXiv:1511.05644,

    Pith/arXiv arXiv

  2. [6]

    Inferring stochastic low-rank recurrent neural networks from neural data.arXiv preprint arXiv:2406.16749,

    Matthijs Pals, A Erdem Sa ˘gtekin, Felix Pei, Manuel Gloeckler, and Jakob H Macke. Inferring stochastic low-rank recurrent neural networks from neural data.arXiv preprint arXiv:2406.16749,

    Pith/arXiv arXiv

  3. [7]

    This supports our choice ofβ= 20in our experiment that has good reconstruction, disentangled latent estimation, and parameter estimation. 101 102 beta 0.70 0.75 0.80 0.85reconstruction R2 101 102 beta 0.10 0.15 0.20partial correlation 101 102 beta 0.7 0.8 0.9latent R2 101 102 beta 0.7 0.8connectivity correlation Figure 6: Metrics w.r.t. the PC penaltyβin ...

    1925

  4. [2013]

    Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Diederik P Kingma. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980,

    Pith/arXiv arXiv

  5. [2018]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    Diederik P Kingma. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114,

    Pith/arXiv arXiv

  6. [2023]

    A revisit of total correlation in disentangled variational auto-encoder with partial disentanglement.arXiv preprint arXiv:2502.02279,

    Chengrui Li, Yunmiao Wang, Yule Wang, Weihan Li, Dieter Jaeger, and Anqi Wu. A revisit of total correlation in disentangled variational auto-encoder with partial disentanglement.arXiv preprint arXiv:2502.02279,

    Pith/arXiv arXiv

  7. [2025]

    One-hot generalized linear model for switching brain state discovery.arXiv preprint arXiv:2310.15263,

    Chengrui Li, Soon Ho Kim, Chris Rodgers, Hannah Choi, and Anqi Wu. One-hot generalized linear model for switching brain state discovery.arXiv preprint arXiv:2310.15263,

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.