- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper introduces PMNLV, a Poisson matrix-normal latent variable model for population-level overdispersion. Binned spike counts are Poisson with a rate equal to a per-neuron tuning term plus a latent gain passed through a softplus link, where the gain has a matrix-normal prior with Kronecker-factored covariance (neuron factor , time factor ). Because the marginal likelihood is intractable, the authors derive a variational EM procedure and then a precision-corrected variant (PC-VEM), motivated by the observation that the optimal Gaussian posterior precision is the prior precision plus a diagonal term, so the target lies outside the Kronecker family; they replace the diagonal by its scalar average and treat as a free parameter of the objective. On synthetic data the method recovers tuning curves and covariance factors and is competitive on co-smoothing; on Allen Visual Coding Neuropixels data it reports near-flat single-neuron variance along the visual hierarchy, shared co-variability peaking in VISp, and a spectral misalignment between the activity covariance and the estimated gain covariance.
Critical strengths
- The generative formulation is well posed and useful. Making the shared gain a matrix-normal latent with explicit neuron and time factors provides an interpretable, scalable handle on population overdispersion that existing single-neuron or low-rank-noise models do not give, and it directly enables the neuron-versus-time decomposition that the results exploit.
- The real-data finding is concrete and biologically interpretable. The claim that gain covariance is spectrally misaligned from the activity covariance, if robust, would be a substantive empirical contribution, since it suggests that the drivers of shared gain differ from the traditionally studied activity covariance.
- The observation behind the correction is a genuine insight. Characterizing the optimal Gaussian posterior precision as prior precision plus a data-dependent diagonal is a clean way to show why the Kronecker family cannot reach the target posterior.
Critical weaknesses
- The precision-correction narrative overreaches and is under-justified. The theoretical analysis is used to motivate a heuristic that changes the objective, and no guarantee or clean characterization of what PC-VEM optimizes is given, yet the framing implies principled consistency. The averaging of a position-, neuron- and time-dependent curvature into a single scalar discards the dependence structure the paper argues is important, and is not defended.
- The real-data evidence is limited and weakly controlled. Results rest on three sessions, small per-region unit counts, and largely two-condition comparisons; the hierarchy trend, the alignment-index result and the multiple comparisons are not supported by uncertainty quantification adequate for the strength of the conclusions.
- The synthetic comparison is not always fair. The out-of-distribution claim for the PLDS-simulated data reports a fraction of oracle for PC-VEM without reporting the corresponding fractions for the baselines, and the co-smoothing evaluation varies the observation model across methods in ways that make direct comparison difficult.
Questions for authors
- Since PC-VEM replaces the true posterior precision with , what objective is being optimized? Is there a valid bound or a fixed-point/consistency argument, or is this best understood as a heuristic estimator?
- How sensitive are , , the covariance-error metrics and the misalignment result to the scalar averaging of ? Please report a variant in which varies by neuron or time, or where a low-rank correction is used.
- In the synthetic comparison, what are the baseline fractions of oracle on the PLDS data? Reporting them alongside PC-VEM's 0.97 is needed for a fair statement of the advantage. Or, specifically, the paragraph starting at line 361 is a bit unclear to me. The authors should clarify or explain these numbers in more detail.
2 comments
Response to Weakness 2:
We agree that the current real-data evidence is relatively limited. To address this weakness, we extend our analysis to an additional 9 sessions from the Brain Observatory component of the Allen Visual Coding database matching the stimuli originally analyzed. These 9 sessions contain an average of 58.0 +/- 19.3 (mean +/- SD) neurons in LGd, 48.7 +/- 11.3 neurons in VISp, 45.9 +/- 10.0 neurons in VISam, and 30.9 +/- 14.1 neurons in VISrl, with 15 presentations per stimulus condition (0°, 45°, 90°, and 135° drifting gratings) at 80% contrast. Applying the same analysis, we find that the results are consistent between the original 3 sessions and the additional 9 sessions (see table below). In both, the marginal variance across regions remains relatively flat while the mean absolute gain correlation starts low in LGd, peaks in VISp, and decreases in VISrl and VISam. We also continue to observe misalignment between activity covariance and gain covariance across all 4 regions.
| Region | Marginal Variance (Median [Q1, Q3]) | Mean Absolute Gain Correlation (Median [Q1, Q3]) | Activity Covariance--Gain Covariance Alignment (Mean ± SEM) |
|---|---|---|---|
| LGd | 0.561 [0.402, 0.763] | 0.260 [0.224, 0.306] | 0.513 ± 0.013 |
| VISp | 0.469 [0.247, 0.808] | 0.440 [0.409, 0.472] | 0.548 ± 0.010 |
| VISrl | 0.411 [0.242, 0.676] | 0.367 [0.328, 0.401] | 0.608 ± 0.017 |
| VISam | 0.418 [0.246, 0.777] | 0.362 [0.320, 0.462] | 0.495 ± 0.016 |
Response to Weakness 3:
This is a valid point to make an OOD claim. The oracle fraction was PC-VEM baseline which favours our method. We have corrected this using the co-smoothing metric described in the paper compared to a Poisson baseline. The methods cannot be directly compared between observation models but the methods can be evaluated by their ability to report the variability in the data which they all capture with different assumptions. In order to keep the evaluation fair, the observation models were kept to their original form.
Response to Question 2:
The covariance errors did not change. For \hat{U}, the relative Frobenius error was 0.320 ± 0.004 under PC-VEM and 0.318 ± 0.004 under the per-neuron variant (mean ± SEM over 12 draws). For \hat{V}, it was 0.248 ± 0.015 and 0.249 ± 0.015. Both differences are smaller than the standard error across draws. In this test, using one number in place of D_{n,t} doesn’t affect how well either covariance factor is recovered. The misalignment result is computed from \hat{U}, whose error did not change. We will add this comparison to the appendix. We additionally ran a variant in which D_{n,t} varies. Each neuron gets its own weight, equal to its average Fisher information over time bins and trials, scaled to have mean one. As in PC-VEM, a single overall scale is then chosen by maximizing the ELBO. We fit the variant to the (LR, M) regime of Table 1, on the same 12 data draws, finding a negligible difference
| Method | Mean | SEM | Mean | SEM |
|---|---|---|---|---|
| PC-VEM | 0.320 | 0.004 | 0.248 | 0.015 |
| per-neuron | 0.318 | 0.004 | 0.249 | 0.015 |
Response to Question 3:
We fit each method with its own observation model in order to preserve the method comparison. We have corrected this by using the co-smoothing metric described in the paper compared to a Poisson baseline given by each neuron’s average spike count in each time bin over the training trials (nats per observation). The comparison doesn’t depend on the models since each method predicts firing rate for the same held-out neurons and all predictions are scored by the same Poisson LL to the same baseline (see table below). Accordingly the paragraph starting at line 361 will be updated.
| Scenario | PC-VEM | GPFA | PGPFA | SNP-GPFA | GP-LVM |
|---|---|---|---|---|---|
| LR LR | 0.0459 | 0.0135 | -0.0228 | 0.0083 | 0.0243 |
| LR Matern | 0.0754 | 0.0285 | -0.0146 | 0.0620 | 0.0418 |
| Matern Matern | 0.0916 | 0.0213 | -0.0169 | 0.0721 | 0.0368 |
| snp | 0.0825 | 0.0411 | 0.0168 | 0.0821 | 0.0547 |
| plds | 0.1738 | 0.0766 | 0.0265 | 0.1443 | 0.1111 |
Response to Weakness 1 (narrative) and Question 1:
Thank you. Section 3.3 will state the E-step constraint set as its own numbered equation, together with both steps. We apply the correction to the constraint set of the E-step. The ELBO, eq. (3.3), is unchanged. What changes is the set of Gaussian posteriors, which in PC-VEM is built from the current neuron and time covariances. Lemma 1 proves that, at every critical point of the ELBO, the posterior precision is the prior precision plus a diagonal. Lemma 2 proves that this diagonal is positive. By the proof of Theorem 1, this form with any positive \gamma is not a Kronecker product of neuron and time covariances. Using one \gamma for the whole diagonal is a heuristic choice.
PC-VEM is an alternating maximization of the ELBO. In the E-step, the constraint set changes only for the variational covariance, by constraining its precision. The posterior covariance is shared across conditions, as in VEM. But its precision is the current prior precision with \gamma added to each diagonal entry. Lemma 1 shows that an optimal Gaussian posterior has precision equal to the prior precision plus a diagonal, and PC-VEM adopts that form with a single \gamma. This observation motivated the change to the optimization constraints and is heuristic. Every positive \gamma gives a positive-definite covariance, and so the ELBO is a lower bound on each iterate by Jensen’s inequality. We do not have a statistical consistency result. For robustness, we will add standard EM algorithm safeguards e.g. step-reduction updates. In all our synthetic experiments the ELBO rose at every iteration, so we expect the safeguard to leave our results unchanged.
Response to Weakness 1 (gamma):
We appreciate this point. It was very helpful, and addressing it has made Section 3.3 clearer. The proof of Theorem 1 uses only that \gamma is positive, so we will restate the theorem for any positive \gamma. The Poisson likelihood contributes one curvature value per neuron and time bin, with nothing linking different entries. All dependence across neurons and time therefore comes from the prior precision. PC-VEM keeps this in its entirety, and the M-step re-estimates every entry of both covariances. A single \gamma does not capture how the curvature varies between entries, and we will state this trade-off in Section 3.3 explicitly. Empirically we also find this important, as in our ground-truth experiments, PC-VEM recovered both covariances more accurately than VEM in every regime.