Pith. sign in

REVIEW 2 major objections 5 minor 14 references

Cross-Subject Modeling for Widefield Calcium Imaging via Atlas-Aligned Spatiotemporal Tokenization

T0 review · 2 major / 5 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read Atlas-aligned tokenization lets one model decode behavior from unseen mice without retraining.

desk verdict Solid multi-subject widefield backbone with real zero-shot continuous decoding; registration dependence is the main soft spot and they already stress-test it. read the letter →

arxiv 2607.09754 v1 pith:IKDJ7LDJ submitted 2026-07-05 cs.CV cs.AIcs.LGq-bio.NC

Mohammad Hosseini , Eray Erturk , Saba Hashemi , Maryam M. Shanechi This is my paper
classification cs.CVcs.AIcs.LGq-bio.NC
keywords widefieldcalciumimagingcross-subjectmodelingatlas-alignedtokenizationmaskedautoencodingzero-shotbehaviordecodingspatiotemporalTransformerfoundationmodelsforneuraldata
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

Widefield calcium imaging movies of cortex are high-dimensional, noisy, and historically modeled one animal or one session at a time, so learned features do not travel. This paper shows that registering every recording to a common brain atlas, chopping frames into fixed spatial patches, and giving every patch a single shared anatomical embedding lets a Transformer pretrain on many mice without any subject- or session-specific knobs. After masked reconstruction pretraining, the frozen embeddings support continuous behavior decoding on entirely new animals and even reconstruction of whole left-out cortical regions. The result is a practical first multi-subject model for this imaging modality and a concrete demonstration that subject-invariant zero-shot behavior decoding is possible when anatomy is used as the common coordinate system.

What carries the argument

Atlas-grounded spatiotemporal tokenization: every frame is registered to the Allen atlas, partitioned into fixed patches, and each patch receives a learnable embedding that is identical across all subjects and sessions; the resulting tokens are processed by a Transformer with joint space-time attention and rotary temporal encoding under a masked autoencoding objective.

What would settle it

Apply controlled large affine misalignments (for example, rotations beyond about 30 degrees plus scale and shift noise) to held-out sessions after pretraining and measure whether zero-shot behavior R-squared falls to the level of chance or of session-specific baselines; if it does not, the claim that atlas-aligned global embeddings are the critical enabler is weakened.

Watch

Extended reading notes

Core claim

A multi-subject model that tokenizes atlas-aligned widefield frames with globally shared spatial embeddings and is pretrained by reconstructing 90 percent of masked spatiotemporal patches learns representations that support zero-shot continuous behavior decoding and left-out region reconstruction on unseen subjects, outperforming single-session and multi-session baselines on two public datasets without requiring session-specific parameters.

Load-bearing premise

That routine affine registration of each session onto a common brain atlas is accurate enough for the shared patch embeddings to stay anatomically meaningful across animals; when that alignment is badly off, zero-shot transfer collapses.

Editorial extensions

If this is right

  • New animals can be decoded for continuous behavior with only a frozen linear or Kalman head, removing the need to retrain session-specific parameters.
  • The same frozen model can reconstruct activity in entirely left-out cortical regions from the remaining brain, giving a zero-shot cross-region prediction tool.
  • Pretraining on one widefield task and dataset transfers under linear probing or light finetuning to a second task and recording setup.
  • Scaling pretraining data volume and Transformer depth continues to improve both finetune and zero-shot decoding, indicating a path toward larger foundation-style models for the modality.

Reading between the lines

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

  • If atlas registration remains the load-bearing step, future multi-lab or freely moving widefield collections will need either better automatic alignment or registration-aware pretraining objectives before zero-shot claims can be trusted at scale.
  • The same global-embedding idea may transfer to other optical modalities that already share a common anatomical reference, such as simultaneous widefield and fMRI or multi-plane two-photon fields of view.
  • Because the decoder is deliberately lightweight and frozen, richer dynamical heads or partial backbone finetuning are natural next levers once larger multi-task corpora exist.
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.

Referee Report

2 major / 5 minor

Summary. The paper introduces WiCAT, a multi-subject model for widefield calcium imaging that combines Allen CCF atlas-aligned spatiotemporal tokenization, globally shared spatial embeddings (no session/subject-specific parameters), a joint space–time Transformer, and aggressive 90% masked autoencoding pretraining. Pretrained on large multi-subject data (Musall and Kondo; 38 subjects, 378 sessions), the frozen backbone supports linear probing, few-shot adaptation, cross-dataset transfer, zero-shot continuous behavior decoding on held-out subjects, and zero-shot reconstruction of left-out cortical regions. Across disjoint pretrain/finetune/zero-shot subject splits, WiCAT outperforms single-session baselines, multi-session SBIND variants that share the same atlas alignment, and adapted NDT2/CEBRA baselines (Table 1; §4.1–4.4), with ablations isolating the roles of global embeddings, MAE vs JEPA, and multi-subject pretraining.

Significance. If the results hold, this is a clear first demonstration of multi-subject pretrained modeling for widefield calcium imaging that achieves subject-invariant zero-shot continuous behavior decoding without session-specific parameters—a capability the authors correctly note has remained elusive even in other neural modalities. The design is carefully controlled (three-way subject splits, frozen-backbone probing, multi-session baselines sharing atlas alignment, chance permutations, decoder-fairness checks, registration perturbation analysis). Strengths include public code/weights, extensive ablations (no pos enc, session-specific, random init, JEPA, masking ratio), scaling curves, and zero-shot neural reconstruction. Absolute zero-shot R² on Kondo is modest, but the transfer argument itself is well supported and advances foundation-style modeling for this modality.

major comments (2)
  1. The central zero-shot transfer claim depends on per-session affine registration to the Allen CCF producing a sufficiently accurate shared coordinate system for global patch embeddings (§3.1). Appendix Table 12 shows that moderate affine noise is tolerated but larger rotations sharply degrade zero-shot R² (e.g., Musall 30°: zero-shot drops from ~0.33 to ~0.13). The Discussion acknowledges this; the manuscript would be stronger if the main text more explicitly quantified residual registration error on the public datasets and stated the operating range of alignment quality under which the zero-shot claim is expected to hold, rather than leaving the dependence primarily in the appendix.
  2. Absolute zero-shot continuous decoding performance remains modest on Kondo (linear probing R² ≈ 0.184; Kalman smoother ≈ 0.287; Table 1–2), even though it significantly exceeds baselines and chance (Table 10). Given that the abstract and §4.2 emphasize “robust” zero-shot decoding, the main text should more clearly contextualize these absolute levels (task difficulty, session variability in Fig. 15) so that “robust” is not over-read as near-ceiling performance. This is a framing issue, not an internal inconsistency in the transfer design.
minor comments (5)
  1. Clarify in the main text (near Table 1 or §4.1) that multi-session SBIND variants use the same Allen atlas alignment as WiCAT, so the comparison isolates tokenization/embeddings rather than alignment alone; this is stated but easy to miss.
  2. Figure 1 caption notes a coarser 2×2 grid for visualization while the model uses 32×32 patches; a brief reminder in the methods that N=16 patches per frame is the default would reduce confusion.
  3. Table 4 vs Table 9: note more prominently that ROI-level averages are not simple averages of subregion MSEs, as already stated in the Table 9 caption.
  4. A few typos and long sentences in the Introduction and Discussion could be tightened for readability (e.g., “Doing so, we establish…”).
  5. Code and weights are promised at the GitHub link; ensuring the release includes the exact subject splits and preprocessing scripts used for Tables 1–4 would maximize reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical multi-subject transfer claims are measured on held-out subjects with external R2/MSE metrics, not forced by definition or self-citation.

full rationale

WiCAT's central claims are empirical ML results: atlas-aligned tokenization plus 90% MAE pretraining without session/subject-specific parameters yields frozen representations that support zero-shot continuous behavior decoding and left-out region reconstruction on subjects never used in pretraining or decoder training (three-way subject split; Table 1; §4.2–4.4). Behavior R2 and reconstruction MSE are external evaluation metrics, not algebraic restatements of fitted parameters. The MAE objective reconstructs masked neural patches; behavior decoding is a separate downstream probe. Ablations (no pos enc, session-specific embeddings, random init, JEPA), chance permutation baselines (Tables 10–11), decoder-fairness controls (Table 8), and atlas-perturbation stress tests (Table 12) provide independent checks rather than tautologies. Self-citations (e.g., SBIND as a single-session baseline) are comparative baselines, not load-bearing uniqueness theorems or definitional premises of the zero-shot claim. No self-definitional equations, fitted-input-as-prediction, uniqueness import, or ansatz-smuggling chain reduces the claimed transfer result to its inputs by construction.

Assumptions & free parameters 6 free parameters · 4 assumptions · 2 invented entities

The central transfer claim rests on domain assumptions about anatomical consistency and atlas registration, plus many standard ML hyperparameters chosen for peak validation decoding. No new physical entities are postulated; the invented pieces are architectural (global atlas patch embeddings, WiCAT MAE pipeline). Free parameters are ordinary training choices, not hidden constants that define the reported R².

free parameters (6)
  • spatiotemporal masking ratio
    Chosen at 90% via validation decoding (Fig. 10); strongly affects pretraining difficulty and reported peak performance.
  • spatial patch size P
    Default P=32 (16 patches/frame); finer P=16 used only for embedding visualization—tokenization granularity is a design choice not derived.
  • Transformer depth / width
    8 layers, 8 heads, d=512 (~36M params) selected after capacity sweeps (Fig. 9).
  • pretraining learning rate and schedule
    Peak LR 6.25e-4, warmup, exponential decay, weight-decay ramp—standard but fitted to training stability.
  • Kalman latent dimension dx
    dx=16 for dynamical decoder experiments (Appendix F.2); improves zero-shot R² but is a free decoder hyperparameter.
  • subject split assignment
    Pretrain/finetune/zero-shot subject partitions (Table 5) are fixed design choices that define the generalization claim.
assumptions (4)
  • domain assumption Large-scale cortical spatial organization is sufficiently consistent across mice that a shared atlas coordinate system can support subject-invariant embeddings.
    Stated in Introduction and Methods as the motivation for atlas-grounded tokenization without session-specific parameters.
  • domain assumption Per-session affine registration to Allen CCF plus resampling to 128×128 yields a usable common spatial frame across datasets and labs.
    §3.1 and Appendix A; load-bearing for global spatial embeddings; stress-tested only with synthetic affine perturbations.
  • ad hoc to paper Aggressive masked reconstruction of neural patches (without behavior labels or subject IDs) yields representations that retain behaviorally relevant dynamics.
    Core pretraining hypothesis; supported empirically vs JEPA and random init but not guaranteed a priori for widefield data.
  • domain assumption Standard Transformer self-attention with RoPE over flattened space–time tokens is an adequate inductive bias for cortex-wide dynamics.
    Architectural choice in §3.1; common in sequence modeling, not proven optimal for this modality.
invented entities (2)
  • WiCAT atlas-aligned spatiotemporal tokenizer with global spatial embeddings
    purpose: Map multi-subject widefield videos into a shared token space without session-specific parameters.
    Architectural construct; independent evidence is only the paper’s transfer experiments, not an external measurement of the embeddings as biological objects.
  • WiCAT MAE pretraining pipeline for widefield calcium videos
    purpose: Learn shared spatiotemporal representations via 90% token masking and lightweight decoder reconstruction.
    Method package combining known MAE ideas with atlas tokens; falsifiable only via replication on held-out data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Cross-Subject Modeling for Widefield Calcium Imaging via Atlas-Aligned Spatiotemporal Tokenization." pith.science (2026). https://pith.science/paper/IKDJ7LDJ

@misc{pith2026260709754,
  author       = {Pith},
  title        = {Pith review of: Cross-Subject Modeling for Widefield Calcium Imaging via Atlas-Aligned Spatiotemporal Tokenization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IKDJ7LDJ}},
  note         = {Machine review of arXiv:2607.09754}
}
read the original abstract

Large-scale, multi-subject widefield calcium imaging provides unprecedented access to brain-wide cortical dynamics. However, the high dimensionality, complex spatiotemporal structure, and substantial task-irrelevant activity in widefield recordings have largely restricted modeling efforts to single-session analyses, limiting scalability and generalization. While multi-subject pretrained models have been explored for some neural modalities, multi-subject models for widefield calcium imaging have not yet been demonstrated; further, subject-invariant zero-shot behavior decoding remains elusive for multi-subject models across neural modalities more broadly. As a first step toward foundation modeling of widefield data, we introduce WiCAT, a multi-subject model that leverages self-supervised pretraining to both outperform single-session models and enable zero-shot behavior decoding on unseen subjects. WiCAT introduces an atlas-grounded tokenization scheme without session-specific components and learns globally shared spatiotemporal representations. Across multiple widefield datasets, the pretrained model supports lightweight downstream decoding, transfers across subjects, tasks, and datasets, and outperforms baseline models. Notably, the model also achieves robust zero-shot continuous behavior decoding and left-out brain region reconstruction on unseen subjects.

Figures

Figures reproduced from arXiv: 2607.09754 by the authors.

Figure 1
Figure 1. Overview of WiCAT architecture. Widefield calcium imaging frames are first registered to a common atlas (Allen Brain Atlas) and reshaped to a common spatial resolution of 128 × 128 to align cortical regions across subjects. Tokenization and Embedding: The aligned recordings are then partitioned into spatiotemporal patches. Each patch is projected into a d-dimensional latent space using a shared convolutional embedde… view at source ↗
Figure 2
Figure 2. Self-supervised pretraining via masked autoencoding. Atlas-aligned widefield recordings are patchified into spatiotemporal tokens and a large fraction of tokens (90%) are masked across space and time. Unmasked tokens are embedded with a shared convolutional embedder and added to the global spatial embeddings before being processed by the Transformer encoder. Masked tokens are replaced by a learnable mask token, info… view at source ↗
Figure 3
Figure 3. Few-shot adaptation. Behavior decoding results are shown for (a) Kondo and (b) Musall datasets for different num￾bers of trials used for adaptation. WiCAT reaches peak perfor￾mance with fewer adaptation trials compared to WiCAT Session￾specific. Each trial consists of 7.00 seconds (Kondo) or 6.83 seconds (Musall) of widefield image frames. comparing against models pretrained directly on the target dataset [PITH_FUL… view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Zero-shot neural reconstruction shows that our model effectively infers the activity of unseen functional areas. Rep￾resentative activity traces from four major cortical regions (MO, SS, VIS, RSP) for a held-out subject. The model’s reconstructed neural dynamics in the…
Figure 5
Figure 5. Figure 5: t-SNE visualization of learned patch embeddings. (a) Each point represents the mean embedding of one spatial patch for one subject, averaged over time and trials. Colors correspond to atlas-aligned patch location, and markers denote subjects. Patches from the same spat…
Figure 6
Figure 6. Figure 6: Performance scales with pretraining data. Behavior decoding performance (R 2± SEM, aggregated across 5 finetuning seeds and all sessions) for finetune and zero-shot sets on (a) Musall and (b) Kondo datasets as a function of pretraining data ratio. similarly results in …
Figure 7
Figure 7. Figure 7: Schematic illustration of the disjoint dataset splits and their roles in training and evaluation. For each split, lighter colors indicate the test set, while darker colors indicate the training and validation data used for model optimization. The pretrain set is used e…
Figure 8
Figure 8. Figure 8: Cortical areas based on the Allen CCF. The map highlights anatomical regions with boundaries. Region correspondences (left hemisphere) are: 1: Primary auditory area (AUDp); 2: Secondary auditory area (AUDs); 3: Primary motor area (MOp); 4: Secondary motor area (MOs); 5…
Figure 9
Figure 9. Figure 9: Effect of model capacity on downstream behavior decoding performance. Behavior decoding R 2 is shown as a function of the number of Transformer encoder layers for the Musall and Kondo datasets. Results are averaged over 3 finetuning seeds and all sessions (mean ± SEM).…
Figure 10
Figure 10. Figure 10: Effect of masking ratio during self-supervised pretraining. Downstream behavior decoding R 2 is shown for different masking ratios on the Musall and Kondo averaged over 3 finetuning seeds and all sessions (mean ± SEM). Performance peaks at a 90% masking ratio, indicat…
Figure 11
Figure 11. Figure 11: Per-pixel R 2 maps for left-out region prediction. Per-pixel R 2 for (a) held-in subjects (top row; pretraining set, sessions seen during pretraining) and (b) held-out subjects (bottom row; finetune and zero-shot sets, with no finetuning performed for this reconstruct…
Figure 12
Figure 12. Figure 12: Qualitative snapshots of neural reconstruction with masked regions. Snapshots of (a) widefield activity (top row) across multiple time frames compared to (b) model predictions (bottom row). In these examples, all visual area (VIS) patches are masked (as indicated by t…
Figure 13
Figure 13. Figure 13: Temporal prediction without future context. Mean squared error (MSE) across all pixels for predicting future frames when only the first part of the trial is provided as input (last 100 frames predicted without context), evaluated across held-in (pretraining set) and h…
Figure 14
Figure 14. Figure 14: Correlation structure of learned patch embeddings. Mean pairwise correlations between patch embeddings, aggregated across sessions. Each panel shows, as a heatmap, the correlation of all other patches with a reference patch: (a) 24, (b) 47, and (c) 49. The reference p…
Figure 15
Figure 15. Figure 15: Session-level factors affecting decoding performance on Kondo. (a) R 2 vs. number of trials per session (color indicates success rate), (b) R 2 vs. session success rate, reflecting behavioral engagement, (c) R 2 vs. neural SNR (mean PC trace variance from PCA), and (d…
Figure 16
Figure 16. Figure 16: Per-session behavior decoding performance on the Musall task dataset. R 2 is shown for zero-shot, few-shot (2, 4, 8, 16 trials), and full-shot settings, averaged over 5 finetuning seeds. Decoding performance improves consistently as additional labeled trials are used …
Figure 17
Figure 17. Figure 17: Behavior decoding traces on the Musall dataset. Predicted behavior is shown over 34.1 seconds (5 trials) for zero-shot decoding using the frozen decoder and full-shot finetuned decoder trained on a representative session. Both zero-shot and full-shot predictions captu…
Figure 18
Figure 18. Figure 18: Per-session behavior decoding performance on the Kondo dataset. Figure conventions are the same as in [PITH_FULL_IMAGE:figures/full_fig_p030_18.png]
Figure 19
Figure 19. Figure 19: Behavior decoding traces on the Kondo dataset. Predicted behavior is shown over 35.0 seconds (5 trials) for zero-shot decoding and full-shot linear decoder trained on a representative session. Figure conventions are the same as in [PITH_FULL_IMAGE:figures/full_fig_p0…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

14 extracted references · 4 canonical work pages

  1. [1]

    org/CorpusID:255999752

    URL https://api.semanticscholar. org/CorpusID:255999752. Azabou, M., Arora, V ., Ganesh, V ., Mao, X., Nachimuthu, S. B., Mendelson, M. J., Richards, B. A., Perich, M. G., Lajoie, G., and Dyer, E. L. A unified, scalable frame- work for neural population decoding. InThirty-seventh Conference on Neural Information Processing Systems,

  2. [2]

    Azabou, M., Pan, K

    URL https://openreview.net/forum? id=sw2Y0sirtM. Azabou, M., Pan, K. X., Arora, V ., Knight, I. J., Dyer, E. L., and Richards, B. A. Multi-session, multi-task neural decoding from distinct cell-types and brain regions. In The Thirteenth International Conference on Learning Representations, 2025. URL https://openreview. net/forum?id=IuU0wcO0mo. Bardes, A.,...

    arXiv 2025

  3. [3]

    cc/paper_files/paper/2019/file/ a10463df69e52e78372b724471434ec9-Paper

    URL https://proceedings.neurips. cc/paper_files/paper/2019/file/ a10463df69e52e78372b724471434ec9-Paper. pdf. Benisty, H., Barson, D., Moberly, A. H., Matsumoto, H., Wu, L., Tseng, H.-Y ., Kaeser, P. S., and Svoboda, K. Rapid fluctuations in functional connectivity of cortical networks encode spontaneous behavior.Nature Neu- roscience, 27:148–158, 2024. U...

    work page doi:10.1038/s41593-023-01498-y 2019

  4. [4]

    org/CorpusID:222349681

    URL https://api.semanticscholar. org/CorpusID:222349681. Chau, G., Wang, C., Talukder, S. J., Subramaniam, V ., Soedarmadji, S., Yue, Y ., Katz, B., and Barbu, A. Pop- ulation transformer: Learning population-level repre- sentations of neural activity. InThe Thirteenth In- ternational Conference on Learning Representations,

  5. [5]

    Churchland, A

    URL https://openreview.net/forum? id=FVuqJt3c4L. Churchland, A. K., Musall, S., Kaufman, M. T., Juavinett, A. L., and Gluf, S. Dataset from: Single-trial neural dynamics are dominated by richly varied movements, October 2019. [Dataset]. Couto, J., Musall, S., Sun, X. R., et al. Chronic, cortex- wide imaging of specific cell populations during behavior. Na...

    2019

  6. [6]

    Erturk, E., Hashemi, S., and Shanechi, M

    URL https://openreview.net/forum? id=jOHgjZaGqd. Erturk, E., Hashemi, S., and Shanechi, M. M. Cross- modal representational knowledge distillation for en- hanced spike-informed LFP modeling. InThe Thirty- ninth Annual Conference on Neural Information Pro- cessing Systems, 2025. URL https://openreview. net/forum?id=hT7Nj7SAQb. Findling, C., Hubert, F., Lab...

    arXiv doi:10.1038/s41592-020-00984-6 2025

  7. [7]

    URL https: //doi.org/10.1038/s41592-018-0109-9

    doi: 10.1038/s41592-018-0109-9. URL https: //doi.org/10.1038/s41592-018-0109-9. Peters, A. J., Fabre, J. M. J., Steinmetz, N. A., et al. Striatal activity topographically reflects cortical activ- ity.Nature, 591:420–425, 2021. doi: 10.1038/ s41586-020-03166-8. URL https://doi.org/10. 1038/s41586-020-03166-8. Radford, A., Kim, J. W., Hallacy, C., Ramesh, A...

    doi:10.1038/s41592-018-0109-9 2021

  8. [8]

    Raut, R., Rosenthal, Z., Wang, X., et al

    URL https://proceedings.mlr.press/ v139/radford21a.html. Raut, R., Rosenthal, Z., Wang, X., et al. Arousal as a universal embedding for spatiotemporal brain dy- namics.Nature, 647:454–461, 2025. doi: 10.1038/ s41586-025-09544-4. URL https://doi.org/10. 1038/s41586-025-09544-4. Ren, C. and Komiyama, T. Wide-field calcium imaging of cortex-wide activity in ...

    2025

Show all 14 references
  1. [9]

    URL https://www.sciencedirect.com/ science/article/pii/S2211124723005387

    doi: https://doi.org/10.1016/j.celrep.2023.112527. URL https://www.sciencedirect.com/ science/article/pii/S2211124723005387. Sharafi, O., Silliman, T., Mokhtari Dowlatabad, H., Niu, G., Walston, S. T., Bouteiller, J.-M. C., Gokoff- ski, K. K., and Lazzi, G. Population-scale an...

    2023 doi

  2. [10]

    URL https: //doi.org/10.1038/s41467-023-44363-z

    doi: 10.1038/s41467-023-44363-z. URL https: //doi.org/10.1038/s41467-023-44363-z. Vahidi, P., Sani, O. G., and Shanechi, M. M. Modeling and dissociation of intrinsic and input-driven neural pop- ulation dynamics underlying behavior.Proceedings of the National Academy of Scienc...

    doi

  3. [11]

    1073/pnas.2212887121

    URL https://www.pnas.org/doi/10. 1073/pnas.2212887121. Vahidi, P., Sani, O. G., and Shanechi, M. BRAID: Input-driven nonlinear dynamical modeling of neural- behavioral data. InThe Thirteenth International Confer- ence on Learning Representations, 2025. URL https: //openreview....

    2025 doi

  4. [12]

    Wang, Y ., Li, C., Li, W., and Wu, A

    URL https://www.sciencedirect.com/ science/article/pii/S0092867420304025. Wang, Y ., Li, C., Li, W., and Wu, A. Exploring behavior- relevant and disentangled neural dynamics with gener- ative diffusion models. InThe Thirty-eighth Annual Conference on Neural Information Process...

  5. [13]

    Ye, J., Collinger, J

    URL https://openreview.net/forum? id=jL0EsbfbAV. Ye, J., Collinger, J. L., Wehbe, L., and Gaunt, R. Neural data transformer 2: Multi-context pretraining for neural spiking activity. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URL https: //openre...

    2023 arXiv

  6. [14]

    14 Cross-Subject Modeling for Widefield Calcium Imaging A

    URL https://openreview.net/forum? id=uyLtEFnpQP. 14 Cross-Subject Modeling for Widefield Calcium Imaging A. Dataset Details We evaluate our model on two publicly available widefield calcium imaging datasets collected from behaving mice: the Musall dataset (Churchland et al., 2...

    2019

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.