acceptodds
Under review as a conference paper at ICLR 2027

Poisson Sparse Component Analysis: A method to parse single-trial neural population activity into factors reflecting distinct neural processes

Abstract

A central aim in systems neuroscience is discovering the distinct neural processes instantiated within the spiking activity of neural populations. Doing so within single trials or unconstrained behaviors presents dual challenges: an analysis method must disentangle the distinct signals mixed within individual neurons, while simultaneously denoising that activity from the stochastic spiking within large populations. To tackle these challenges, we present Poisson Sparse Component Analysis (pSCA), an unsupervised approach for discovering latent factors reflecting distinct neural processes in spiking datasets. pSCA seeks latents that occur sparsely in time and reside in orthogonal subspaces, enabling the decomposition of population activity into signals reflective of distinct neural processes without prior knowledge of these processes’ identity or timing. Through a convolutional autoencoder architecture and masked training procedure, pSCA recovers denoised latent factors that align with identifiable task events and behavioral variables in multiple datasets across mouse and monkey cortex. Without supervision, pSCA finds a decomposition of spiking activity in which simpler components combine to produce more complex behaviors, exposing compositional structures alternative approaches do not yield.

Primary area:
applications to neuroscience & cognitive science
Keywords:
neuroscience, dimensionality reduction, interpretability

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PWLqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes an extension to a recent method, sparse component analysis, such that it can be used on single-trial spike count data without averaging or pre-smoothing. It adapts SCA primarily with a Poisson reconstruction loss (with sparsity and orthogonality regularization). It is compared against standard methods such as GPFA and LFADS, showing that pSCA recovers the groundtruth synthetic latents accurately with much less training time, and behaviorally isolated latents in monkey hand-cycling and mouse free behavior datasets.

Critical strengths
  • overall a very well-written and clearly structured paper. I think the LLM-assisted polishing introduced some oddities (like the repeated usage of "researchers" in the intro), but as a whole unproblematic. Similarly, the figures are of high quality.
  • the paper proposes a fairly straightforward and obvious extension to SCA, and shows that it indeed produces spatiotemporally isolated latents while incurring significantly shorter runtime than LFADS and GPFA
  • it is evaluated on two real datasets, one standard and one relatively novel, and finds latents that correlate well with intuitive aspects of behavior or EMG
Critical weaknesses
  • the proposed extension is relatively incremental over SCA, which is not a problem by itself, and the simplicity and intuitiveness of the proposed method are advantages. But given the limited methodological innovations, the evaluation is then fairly limited in my opinion.

  • Essentially, pSCA finds the correct GT latents in a synthetic task, and produces more behaviorally-isolated latents in the monkey and mouse datasets. It is acknowledged that the monkey dataset is a well-used benchmark with clear expectations for what one should find, nevertheless, showing that latents have selectivity for prep vs. move or direction of cycling doesn't say much about the method. Though it is interesting that it's much better than the other methods, including SCA.

  • Similarly, for the mouse dataset, (cortical) latents are found to be isolated to labeled behavioral motifs, but this is only done in a qualitative fashion.

  • more generally, I think comparison to more recent baseline methods for unsupervised latent variable models in neuroscience may be warranted, including those that do require pre-smoothing, many of which the authors cite. LFADS and GPFA are both now more than 10 years old.

  • I would have liked to see more interpretation of the latents beyond "they correlated to individual behavioral dimensions". For instance, where are the neurons that participate strongly in each latent, what is the spread of loading for each neuron in the first place, is that biologically plausible, what do the learned conv filters look like for the real datasets, how do the results change with latent dimensionality, etc. The methodological innovation and evaluation are a little thin for an ICLR paper, so I would then at least expect to see some level of scientific discovery.

Questions for authors

Technical clarifications:

  • If I understand correctly, the timepoint-wise MLP’s takes as input the spike trains, and its output is then convolved with the learned filter. Are the conv filters of the encoding model causal? Also, what is meant by “the filter weights are learned independently at each window position”? As in their values can change relative to absolute time, or simply that every value in the kernel is free to change flexibly as in a standard learned convolution in deep learning, but subject to l1 normalization?
  • lambda_sparse is adaptive but lambda_orth requires hyperparam tuning? But if I understand correctly, the sparsity term is adaptive in the sense that it’s pinned to the reconstruction loss by a fixed proportionality, which is itself a hyperparameter?
  • how to choose latent dimensionality k?

More broadly, see the last two points in weaknesses, i.e., 1) further and more competitive baselines published in the last 5 years, and 2) scientific interpretation of learned latents / filters / loadings beyond "they correlate well with behavior".

10 comments

Thanks for the feedback and acknowledging that simplicity and intuitiveness can be advantageous. Indeed, our goal of the method is to have something that is easily usable and intuitive for the neuro community. Our paper thus focused on including the four essential model innovations expanding on the original SCA framework (Section 2) that enable the model to work effectively out-of-the-box on several datasets, rather than extra bells-and-whistles.

In terms of the limited evaluation, based on your feedback (and from other reviewers), we have now substantially increased the evaluation of the method in our revision. This includes providing comparisons to more recent methods (CEBRA) and substantial further scientific exploration of the mouse naturalistic arena dataset (including on 4 additional mice), which is detailed in further responses below.

AuthorsOpenReview ↗

Re: Comparisons to more recent baselines.

“More generally, I think comparison to more recent baseline methods for unsupervised latent variable models in neuroscience may be warranted, including those that do require pre-smoothing, many of which the authors cite. LFADS and GPFA are both now more than 10 years old.”

We thank the reviewer for this suggestion, and believe this addition substantively strengthens the manuscript. To include a recent, widely used latent variable model, we fit CEBRA (Schneider et al., 2023) to the cycling dataset. CEBRA uses contrastive learning to pull together neural activity patterns associated with similar behavioral variables or nearby time points, and to push apart those associated with dissimilar behaviors or distant time points. We fit both a fully unsupervised variant (CEBRA-time) – which would be most comparable to pSCA and the other unsupervised comparisons – and also a variant supervised by EMG and kinematics (CEBRA-behavior) (Figs. 2 and A.15).

Of the competing methods we tested, other than SCA, CEBRA is the only one explicitly designed to produce interpretable latent representations. Even so, it did not demix the processes in this dataset. CEBRA-behavior captures externally observable aspects of the task, such as cycling, but fails to isolate internal processes such as movement preparation, and preparation to stop, which pSCA recovers without behavioral supervision (Fig. 2). CEBRA-time demixes poorly as well (Fig. A.15), and both variants predict held-out neurons worse than methods with a reconstruction term in their objective function (pSCA, GPFA, and LFADS). Across all methods, pSCA remains the most effective at balancing predictive performance, training speed, and demixing (Fig. 2). In response to another reviewer, we also applied post-hoc sparsifying transformations (varimax and SCA) to the latents learned by GPFA, LFADS, CEBRA-time, and CEBRA-behavior. Even after these transformations, pSCA's latents remained the most demixed (Figs. A.16–A.18).

We believe all of the CEBRA results highlight the strengths of our approach. We currently show CEBRA-behavior in Fig. 2 and CEBRA-time in Appendix, but have no strong preference about which variant appears in the main text. If you find a particular one more compelling, we are happy to feature it instead.

AuthorsOpenReview ↗

Re: Scientific interpretation of the learned latents.

“I would have liked to see more interpretation of the latents beyond "they correlated to individual behavioral dimensions". For instance, where are the neurons that participate strongly in each latent, what is the spread of loading for each neuron in the first place, is that biologically plausible, what do the learned conv filters look like for the real datasets, how do the results change with latent dimensionality, etc. The methodological innovation and evaluation are a little thin for an ICLR paper, so I would then at least expect to see some level of scientific discovery.”

We thank the reviewer for pushing us on this point. To move beyond correlating latents with behavior, we fit pSCA separately to fM1 and DLS recordings from four additional mice. We elaborate on these results in the main text (new Fig. 4; Section 3.4) and appendix (Figs. A.26, A.28–A.31), and briefly summarize the key findings here:

  1. ) Consistent latents across animals and regions. In four of the five mice, pSCA found an eating-selective latent in both fM1 and DLS that closely tracked muscle activity (Fig. 4A). Because each region of each animal was fit separately, and the striatal recording location differed from mouse to mouse, this recurrence indicates that pSCA recovers structure that generalizes across brain regions and animals.

  2. ) Reproducible inter-regional timing. Identifying the same latent in each region allows its timing to be compared directly. During eating, fM1 leads DLS by an identical 80 ms in three of the five mice (Fig. 4A), whereas at rearing offset a striatal transition latent peaks before the cortical one, also generalizing to held-out bouts (Fig. 4D) - see Fig. A.31 for another striatal transition signal within a different mouse.

  3. ) Biologically plausible loadings. In response to the reviewer’s question about where contributing neurons are located, we quantified the spatial clustering of each eating latent’s loadings along the probe with Moran’s I (Fig. 4B,C; Section A.6.6). In every mouse, striatal eating latents were carried by spatially clustered neurons, whereas cortical ones were distributed across the population, consistent with the convergent organization of corticostriatal projections.

  4. ) A testable hypothesis. Together, these results suggest that cortex leads striatum within behaviors requiring moment-to-moment forelimb control, whereas striatum leads cortex at transitions between behaviors, consistent with the proposed role of striatum in action selection (Markowitz et al., 2018). This hypothesis rests on pSCA identifying the same process independently in each region and animal, something methods with latents co-active across all behaviors, such as GPFA (Fig. 3B), cannot provide.

Reviewer_E7DEOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces pSCA, an unsupervised method for extracting sparse, temporally separated latent factors from neural spike counts. It extends SCA with a Poisson observation model, learned temporal filters, coordinated dropout, and adaptive sparsity regularization. Synthetic experiments and monkey and mouse recordings demonstrate improved factor separation, with reported predictive performance comparable to LFADS on the cycling dataset.

Critical strengths
  1. Jointly learning denoising and sparse factors addresses a practical limitation of applying SCA to noisy single-trial recordings. The synthetic ablations help establish why temporal filtering and coordinated dropout matter.
  2. The applications recover recognizable task-related structure and generate interesting hypotheses about naturalistic behavior. The reported balance between interpretability, prediction accuracy, and training time is promising.
Critical weaknesses
  1. GPFA and LFADS are not designed to produce sparse individual factors. Comparisons with these methods followed by unsupervised SCA or a sparsity-promoting rotation are needed to establish the benefit of joint fitting.
  2. Appendix A.4.2 does not clearly state whether test trials are excluded when fitting the latent models, although it explicitly splits trials for the downstream GLM. This ambiguity affects the claim of generalization to held-out trials.
  3. The naturalistic analysis uses behavior-defined bouts for training, and Figure 3D selects examples that maximize transition selectivity. Results on unsegmented recordings and quantitative evaluation on held-out bouts are needed to establish how representative these findings are.
  4. Decoder orthogonality is encouraged by the objective and applies in log-rate space under the exponential link. It does not independently establish that the inferred biological processes occupy orthogonal subspaces in firing-rate space, as the stopping-preparation interpretation may suggest.
Questions for authors
  1. Were held-out trials excluded from all latent-model fitting and hyperparameter selection in Figure 2E, or only from GLM fitting?
  2. How does pSCA compare with unsupervised sparse decomposition applied to GPFA- or LFADS-denoised activity?
  3. Do the cortical–striatal differences persist on held-out bouts without selecting examples for maximal transition activity?
8 comments

Thank you for your extremely helpful review. We’re really appreciative you found the paper clear and found our balance of interpretability, accuracy, and training time promising. Your suggestions led to several new analyses (detailed below) that we believe substantially strengthen the paper, especially the analyses using post-hoc sparsifying rotations.

AuthorsOpenReview ↗

Re: Comparison with sparse decomposition applied to GPFA- or LFADS-denoised activity.

“GPFA and LFADS are not designed to produce sparse individual factors. Comparisons with these methods followed by unsupervised SCA or a sparsity-promoting rotation are needed to establish the benefit of joint fitting.”

“How does pSCA compare with unsupervised sparse decomposition applied to GPFA- or LFADS-denoised activity?”

We thank the reviewer for this suggestion and have added this comparison to demonstrate the importance of our end-to-end fitting approach (Figs. A.16–A.18).

Background:

Existing methods such as GPFA and LFADS are designed to infer the level of smoothing needed to denoise single-trial spiking data, not find inherently interpretable latents. Interpreting GPFA and LFADS factors after post-hoc sparsifying transformations could serve to further justify pSCA’s joint denoising and demixing training scheme.

Analysis:

On the cycling dataset, we took the latents learned by GPFA and LFADS, and then applied two different post-hoc sparsifying rotations to see if either could yield comparable latents to pSCA. The two posthoc rotations were a varimax rotation, and applying SCA itself. We also applied both rotations to CEBRA-time and CEBRA-behavior latents (which we fit in response to Reviewer PWLq), and compared all of these with pSCA and with SCA applied to pre-smoothed data. We then quantified how well each method’s latents separate preparation from movement, forward from backward cycling, and stopping from movement.

We note that applying SCA post hoc in this way is an off-label use of the method. SCA was designed to reduce the dimensionality of neural activity, not to rotate latents that have already been reduced, and the publicly available codebase does not support this use. SCA sets its sparsity penalty relative to the initial reconstruction loss. When SCA is fit with the same dimensionality as the latents it receives, it can reconstruct its input almost perfectly from the start. The initial reconstruction loss is then practically zero, and so is the resulting penalty. To address the reviewer’s question, we used SCA without the default sparsity penalty, and rather used a fixed sparsity penalty (independent of initial reconstruction loss) that led to a rotation.

Results:

Even after post-hoc sparsification of the other methods’ factors using SCA or varimax, pSCA’s latents remain quantitatively the most demixed for all three pairs of processes analyzed (Fig. A.18). For preparation vs. movement pSCA’s selectivity (~0.46) is well above the best post-hoc result (~0.32, CEBRA + SCA) - an even larger difference is apparent when using pSCA with coupled coordinated dropout (~0.51). For other task epochs, the difference may be somewhat smaller - e.g. for stopping vs. movement, GPFA + SCA comes closest (~0.30 vs. ~0.34 for pSCA) - yet this stepwise approach consistently still falls short across analyzed epochs. Qualitatively, LFADS and GPFA + post-hoc SCA produce reasonable-looking latents (Fig. A.17), but they are less demixed and noisier than pSCA’s. In some cases, the stepwise approach misses structure entirely: GPFA + SCA has no latent corresponding to pSCA’s preparatory latent z9 (Fig. A.17). In sum, if the goal is interpretable latent factors on binned spiking data, there is little reason to fit LFADS or GPFA followed by a sparsifying rotation - not only does pSCA learn more demixed, less noisy latents then LFADS + post-hoc SCA - it also trains roughly 2x faster than GPFA and 10× faster than LFADS alone (0.59 hr vs. 5.82 hr for LFADS; Fig. 2F).

We attribute pSCA’s advantage to fitting everything in one stage. When sparsification follows a separate denoising step, the rotation can only recombine latents that were already learned. If the denoising model has mixed processes across its latents or embedded spurious activity into the latent space, a post-hoc rotation may not fully separate them. pSCA learns the denoising and the sparse decomposition jointly, so the temporal smoothing is shaped around temporally distinct factors rather than fixed beforehand.

AuthorsOpenReview ↗

Re: Held-out trials in Appendix A.4.2.

“Appendix A.4.2 does not clearly state whether test trials are excluded when fitting the latent models, although it explicitly splits trials for the downstream GLM. This ambiguity affects the claim of generalization to held-out trials.”

“Were held-out trials excluded from all latent-model fitting and hyperparameter selection in Figure 2E, or only from GLM fitting?”

We thank the reviewer for pointing out this ambiguity. Our approach uses a standard co-smoothing protocol (Pei et al., 2021; Dabholkar & Barak, 2025), in which the pSCA models are fit on all trials, but only on a subset of neurons. In co-smoothing, on each fold (5 here), the latent model is fit to 80% of the neurons across all trials, and latents are inferred from these held-in neurons only. A Poisson GLM readout is trained to map those latents to the held-out neurons on 80% of the trials. The readout is tested on its prediction of the held-out neurons on the remaining 20% of trials. The held-out neurons on the held-out trials are therefore never seen, either when fitting the latent model or when fitting the readout. Our setup is also stricter than the Neural Latents Benchmark setup (Pei et al, 2021) in one respect in that the latent models never see the held-out neurons at all, not even on training trials.

Crucially, no hyperparameters were tuned using held-out data. pSCA sets its sparsity penalty automatically (section 2.7, section A.4.7), and every other hyperparameter uses the same default value across all datasets. The reported generalization therefore involves no selection using the test set. We revised Appendix A.4.2 to state the protocol explicitly.

Reviewer_1TNSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Poisson Sparse Component Analysis (pSCA), an unsupervised convolutional autoencoder for extracting latent factors from single-trial and naturalistic neural spike-count data.

The authors evaluate pSCA on synthetic data with known latent factors, macaque motor-cortex recordings during hand-cycling, and mouse motor-cortex and striatal recordings during naturalistic behavior. They report recovery of synthetic factors, interpretable factors associated with task epochs in the macaque data, and factors with behavior specific or shared activity in the mouse data.

The paper frames the naturalistic findings as hypotheses for subsequent investigation and notes that pSCA relies on temporal sparsity and approximate subspace orthogonality.

Critical strengths
  1. pSCA addresses an important practical gap between methods that denoise single-trial spike counts and methods designed to produce interpretable, demixed factors.

  2. The evaluation includes both known-ground-truth simulations and neural recordings from distinct settings. The synthetic recovery and ablation experiments help clarify which components contribute to demixing, while the macaque comparison assesses interpretability alongside held-out prediction.

Critical weaknesses
  1. Behavioral selectivity is summarized using the same behavior labels that the paper uses to interpret the latents; this can illustrate correspondence, but does not by itself establish that the factors are distinct neural processes or that the proposed compositional account is correct.

  2. The method’s intended demixing depends on temporal sparsity and approximately orthogonal neural subspaces. These are consequential modeling assumptions: when processes overlap in time or occupy similar subspaces, the decomposition may not match the underlying processes. The paper acknowledges this limitation, but the empirical robustness tests are mainly synthetic, and the real-data analyses do not establish how often these assumptions hold or how sensitive the inferred factors are to deviations from them.

Questions for authors
  1. For the mouse naturalistic dataset, can you quantify the stability of the inferred factors across random initializations, model fits, or recording sessions? Are the behavior-specific and transition-related findings reproducible under these variations?

  2. How sensitive are the key biological interpretations to the number of latent dimensions and to the sparsity and orthogonality settings?

5 comments

Thank you for your careful reading of our work, and we’re very appreciative that you found value in our method and our combination of ground-truth simulations, ablations, and recordings from distinct experimental settings. We detail the additions we have made to the paper below to address your questions and concerns - thanks again for your feedback, it has helped us to strengthen our paper.

AuthorsOpenReview ↗

Re: Stability of the inferred factors

“For the mouse naturalistic dataset, can you quantify the stability of the inferred factors across random initializations, model fits, or recording sessions? Are the behavior-specific and transition-related findings reproducible under these variations?”

We thank the reviewer for this question, and address each source of variation in turn.

Random initializations.

Background. pSCA initializes its encoder and decoder with principal component analysis (PCA), so random weight initialization is not a source of variability. Two stochastic elements remain in training: the coordinated dropout masks, which are resampled on every forward pass, and the random ordering of training batches.

Analysis. To measure their effect, we fit pSCA to the arena dataset with 5 random seeds and regenerated the behavioral-specificity heatmaps from Fig. 3 for each fit. Because latent order is arbitrary, the same “eating” latent may sit in dimension 0 in one fit and dimension 1 in another, so we matched each fit’s latents to those of the original fit by aligning heatmap columns before comparing them.

Result. Across seeds, pSCA recovered nearly identical behavioral specificity and latent time courses (correlation between heatmap columns = 0.92 ± 0.04; new Fig. A.22).

Model fits on held-out data.

Background. If pSCA’s findings depended on the particular bouts used for training, they would not generalize to unseen data.

Analysis. We refit pSCA while first withholding 20% of the bouts of each behavior, and then with the transition bouts themselves withheld. We compared behavior-specificity heatmaps, transition-related activity, and pseudo-R² between training and held-out bouts.

Result. With 20% of bouts withheld, the behavior-specificity heatmaps were largely unchanged (heatmap column correlation = 0.79 for fM1, 0.88 for striatum; new Fig. A.26). With transition bouts withheld, transition-related activity was still largely recovered on those unseen bouts (new Figs. 4D, A.27, A.31). Pseudo-R² was the same on training and held-out bouts, both overall ( Fig. A.28) and for transition-related activity specifically (R2 with held-in latents in Fig. A.27), so the model is not overfitting to the bouts it was trained on. This matches our cycling results, where pSCA’s pseudo-R² on held-out neurons is high relative to other methods (Fig. 2E). Together, these results show that the behavior-specific and transition findings are robust to held-out data.

Recording sessions.

Background. If pSCA’s findings depend on fits to individual mice, it would not find common compositional structure or latents shared across animals. Identical behavioral specificity is not expected across mice, because the unconstrained paradigm lets each mouse perform each behavior at a different frequency, and some mice never perform certain behaviors. Transition structure is also difficult to compare across mice, mainly because the striatal recording location differs in each animal.

Analysis. We fit pSCA to recording sessions from each of four additional mice, and compared behavioral specificity, eating latents, and transition latents across animals.

Result. pSCA recovers largely shared compositional structure across mice, including latents unique to eating and grooming, and latents shared across types of climbing (Figs. A.29). Eating is the most frequent behavior in every mouse, and in forelimb motor cortex pSCA found strikingly similar eating latents in four of the five mice (Fig. 4A). In striatum, pSCA found multiple eating-specific dimensions in every mouse (new Fig. A.29), which may reflect convergent input to dorsolateral striatum from across sensorimotor cortex during the complex coordination of food handling, chewing, and swallowing. Within individual mice, pSCA also consistently found latents that peaked at the onset or offset of specific behaviors. One example, now highlighted in new Fig. 4D, is a transition latent reliably active at rearing offset, which peaks in dorsolateral striatum before fM1 (another shown in A.31). Thus, despite inter-animal differences, pSCA finds consistent compositional structure across animals and consistent transition signals within them.

AuthorsOpenReview ↗

Re: Modeling assumptions and sensitivity to settings

“The method’s intended demixing depends on temporal sparsity and approximately orthogonal neural subspaces. These are consequential modeling assumptions: when processes overlap in time or occupy similar subspaces, the decomposition may not match the underlying processes. The paper acknowledges this limitation, but the empirical robustness tests are mainly synthetic, and the real-data analyses do not establish how often these assumptions hold or how sensitive the inferred factors are to deviations from them.”

“How sensitive are the key biological interpretations to the number of latent dimensions and to the sparsity and orthogonality settings?”

pSCA assumes that distinct computations are separated in time and occupy distinct subspaces. Below, we argue 1) both assumptions are biologically grounded, 2) together they keep pSCA from inventing structure not present in the data, and 3) pSCA’s conclusions are insensitive to the settings that implement them. Cross-animal consistency (new Section 3.4) further shows that this structure is reproducible.

Sparsity

Assumption. Temporal sparsity formalizes the idea that distinct processes happen at distinct times. Many computations are sequential by definition (e.g., motor preparation precedes execution), so sparsity makes “temporally distinct” precise without imposing any temporal order. pSCA requires processes to be only partly disjoint in time (Section 2.3). When a process is not sparse, pSCA does not invent sparse structure, accurately recovering a slow, non-sparse ramp latent in simulation (Fig. A.6). Similarly, pSCA does not invent sparse cycle-specific latents (e.g., active only on the fourth of seven cycles) in primary motor cortex during cycling (Fig. A.15), consistent with prior work showing that M1 carries no information unique to individual cycles (Russo et al., 2020).

Sensitivity. A sparsity penalty that is too strong could over-split a process into multiple sparse factors (Section A.1.2). pSCA avoids this by scaling the penalty throughout training so that its gradient norm matches that of the reconstruction loss (Section 2.7). On the cycling dataset, this adaptive penalty matched the held-out performance of the best of six fixed penalties, while avoiding the collapse seen at λs ≥ 0.1 (Fig. A.10). This automated procedure avoids such sensitivity.

Orthogonality

Assumption. Approximately orthogonal population subspaces are widely observed (e.g., Elsayed et al., 2016; Kobak et al., 2016; Ames & Churchland, 2019; Inagaki et al., 2022), and are thought to let populations maintain multiple processes without interference. Just as importantly, and as we elaborate on in our new section A.1.2 – pSCA’s collinearity penalty also guards against false positives from the sparsity penalty. Without it, pSCA fit with more factors than the data contain can divide activity into more factors than actually exist. Such over-splitting requires highly collinear factors, and our ablation confirms it: without the penalty, pSCA populates a spurious extra dimension (Section A.3.3).

Sensitivity. The orthogonality setting matters in the opposite direction from sparsity - an orthogonality penalty that is too weak risks over-splitting. We set its weight to 0.1% of the initial reconstruction loss in all experiments, so it scales with each dataset (Section 2.5). At this setting, pSCA learns approximately orthogonal loadings in all datasets without harming reconstruction, recovers the correct dimensionality in simulation (Fig. 1), and finds no cycle-specific factors in the cycling data (Fig. A.15). Therefore a stronger penalty would only enforce structure the loadings already have. Users therefore need no trial-and-error tuning. This default already achieves the approximately orthogonal loadings our hypotheses rely on.

Dimensionality

Sensitivity. Together, sparsity and the collinearity penalty make pSCA robust to over-specifying the number of latent dimensions, since extra dimensions are generally zeroed out. In simulation, pSCA recovers the correct dimensionality even when fit with surplus dimensions, including under violations of its assumptions (Figs. 1, A.6). On the cycling dataset, doubling the number of latents from 20 to 40 leaves the interpretation unchanged. pSCA still recovers dimensions uniquely active during preparation, forward and backward cycling, and stopping (Fig. A.14).

Summary

pSCA’s assumptions make it robust to false positives. It does not invent sparse structure where none exists, and it zeroes out surplus dimensions rather than filling them with spurious factors, all with default settings that need no per-dataset adjustment. The cost is possible false negatives. Sparse processes in highly overlapping subspaces may not be separated. For a method meant for scientific discovery, where spurious structure does more harm than conservatively missed structure, we believe this is the right trade-off.

Comments on the paper

1 comment

We thank all the reviewers for their positive remarks and very constructive feedback that has enabled us to substantially strengthen the manuscript. We summarize our many additional analyses below, which are expanded upon in the individual reviewer responses:

  1. Stability and held-out generalization (Reviewers: 1TNS, E7DE)

    Background: Reviewers asked whether the inferred factors are stable, and whether results generalize to held-out behavioral bouts.

    Analysis: We refit pSCA with five random seeds and with bouts withheld from training, and clarified that in co-smoothing (cross-validation) evaluation protocol, no held-out data was used for training or hyperparameter selection.

    Result: Factors are consistent across seeds (Fig. A.22). Scientific results (Behavioral specificity (Fig. A.26) and transition latents (Figs. 4D, A.27)), and reconstruction performance (Fig. A.28) are preserved on held-out bouts, demonstrating scientific robustness.

  2. Additional mice and scientific interpretation (Reviewers: 1TNS, PWLq)

    Background: Reviewers asked whether pSCA’s findings generalize beyond a single animal, and for scientific interpretation beyond correlation with behavior.

    Analysis: We fit pSCA to recordings from four additional mice and compared latents, inter-regional timing, and loadings across animals and regions.

    Result: pSCA identifies the same eating latent independently in forelimb motor cortex (fM1) and dorsolateral striatum (DLS) in four of five mice (new Figs. 4, A.29). During eating, muscle-related latent activity in fM1 leads DLS by an identical 80 ms in three mice (Fig. 4A), while in one mouse DLS consistently leads fM1 at rearing offset (Fig. 4D). Another finding is that DLS eating-related latents are carried by spatially clustered neurons, unlike in fM1 (Fig. 4B,C). Together, these results suggest testable hypotheses about how the two regions contribute to behavior, which are now elaborated on in the new main text section 3.4.

  3. Modeling assumptions, theory, and clarifications (Reviewers: 1TNS, E7DE, PWLq)

    Background: Reviewers asked how pSCA’s sparsity and orthogonality assumptions affect its results, and raised concerns regarding whether its loading penalty supports claims about neural geometry.

    Analysis: We added an appendix section on the theory behind how orthogonality prevents over-splitting of latents. We characterized sensitivity to latent dimensionality, and added a model-free alignment-index analysis of the neural activity.

    Result: The sparsity and collinearity (orthogonality) penalties used together prevent over-splitting (section A.1.2, Figs. A.2 and A.6), and both are set without tuning. Results are robust to over-specifying the dimensionality (Figs. 1, A.14), which the orthogonality penalty is particularly important for. We also further confirmed the distinctness of processes found in the cycling dataset using a model-free analysis that illustrates that preparing to stop does not reuse the subspace of preparing to start (Fig. A.12).

  4. New baselines and post-hoc sparsification (Reviewers: E7DE, PWLq)

    Background: Reviewers asked for more recent baselines, and also how pSCA compared to a step-wise approach of taking alternative methods’ latents followed by post-hoc sparsifying rotations.

    Analysis: We added an additional baseline comparison of CEBRA (time- and behavior-supervised), and applied post-hoc varimax rotations and SCA to the latents of GPFA, LFADS, and CEBRA.

    Result: With the addition of CEBRA, pSCA remains the strongest, most efficient tool for denoising and demixing distinct neural processes (Figs. 2, A.15, A.19). Even after applying sparsifying rotations to factors from other methods, pSCA remains the most consistent and quantitatively effective method for finding denoised and demixed latent factors, all while training substantially faster than stepwise denoising then demixing approaches (Figs. A.16, A.17, A.18) .

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.