acceptodds
Under review as a conference paper at ICLR 2027

Spherical Recognition-Parameterized State Space Models for Learning Shared Neural and Behavioral Dynamics

Abstract

Simultaneous recordings of neural activity and behavior offer a window into how brain dynamics relate to an animal’s actions. However, recovering shared dynamical structure from multimodal timeseries is challenging when each modality contains substantial private variation. Generative approaches must reconstruct this variation, while standard contrastive methods learn shared representations without explicitly modeling their evolution. We introduce sRP-SSM, a spherical recognition-parameterized state space model that jointly learns cross-modal representations and stochastic latent dynamics without reconstructing observations. The spherical geometry allows recognition factors to remain log-linear in the latent state even with observation-dependent confidence, so evidence from each modality and from past timesteps combines by simple addition. We derive a scalable contrastive training objective that approximates the model’s evidence lower bound, connecting contrastive representation learning with probabilistic modeling of shared dynamics. In synthetic experiments, sRP-SSM combines strong shared-state recovery and cross-modal retrieval with more accurate dynamics recovery than competing generative and contrastive methods. In simultaneous neural and 3D pose recordings from a freely moving mouse, sRP-SSM improves cross-modal retrieval, future prediction, and keypoint decoding. During rearing, it detects behavior from neural activity and recovers bout-averaged neural spectral patterns from pose. These results show how using shared dynamics to guide representation learning can improve the recovery of neural–behavioral relationships, establishing a foundation for dynamics-informed learning across complementary modalities.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
recognition-parameterized models, timeseries, multimodal, product of experts, neuroscience, variational inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_riuZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper extends recognition-parameterized models (RPMs) to multimodal timeseries in order to learn shared latent dynamics without reconstructing observations. A shared latent evolves on the sphere: modality-specific encoders map observations to von Mises-Fisher recognition potentials, evidence is fused by adding natural parameters (a product-of-experts / assumed-density filter whose gain discounts past evidence), and the latent dynamics are a learned orthogonal rotation plus vMF diffusion. Because vMF factors remain log-linear in the latent even with observation-dependent concentration, fusion and contrastive logits stay analytic. The authors derive a CLIP-like minibatch objective argued to approximate the RPM ELBO, consisting of a candidate-based recognition reward plus an analytic KL to the transition prior. In synthetic experiments the minibatch objective tracks a reference ELBO estimate and the model gives the lowest dynamics-recovery error, lying near an oracle Pareto frontier. On mouse LFP plus 3D pose it improves cross-modal retrieval and LFP-to-keypoint decoding, and yields rearing-related structure.

Critical strengths
  • The model is principled and unifies contrastive representation learning with explicit shared dynamics. The vMF construction simultaneously supports additive evidence fusion, an analytic transition KL and a uniform stationary prior, giving an elegant answer to how to infer one latent state from either or both modalities without an observation model. This is consequential because it cleanly separates reconstruction-driven private-variation bias from the shared-dynamics objective.
  • The paper gives a convincing mechanistic diagnosis of when joint learning matters. The demonstration that static contrastive alignment recovers instantaneous structure but loses weakly informative yet persistently predictive signal, and that the no-transition ablation fails only when the recurrent filter is also removed, provides a concrete account of the benefit of dynamics-informed representation rather than a leaderboard gain.
  • The spherical-geometry argument is well developed. Appendix D.1 carefully shows why a first-order critic requires constant precision in the Euclidean case and how fixed radius removes the radial curvature, which is a genuinely useful conceptual contribution.
Critical weaknesses
  • The probabilistic justification is weaker than the framing. A finite, non-decomposable minibatch objective with a candidate-estimated normalizer does not by itself give a valid ELBO; the claim that training optimizes an approximation to the ELBO needs clearer caveats and, ideally, an analysis of the approximation gap.
  • The empirical advantage is small and the strongest baseline is often close. In retrieval, sRP-SSM (35%) versus CLIP-posthoc (33%) is within the reported error; for keypoint-to-LFP decoding CLIP-posthoc actually wins (R2R^2 0.166 versus 0.124). With a single animal and no cross-animal validation, the practical case for joint dynamics is not established.
  • The ablation weakens the role of the stochastic dynamics. Since the transition term and backpropagation through the filter are largely redundant, and the no-transition model matches the reference when the filter is present, it is unclear how much of the benefit comes from the explicit transition model rather than from the recurrent inductive bias.
Questions for authors
  • Can you bound or at least empirically quantify the gap between the minibatch training objective (finite candidates and a minibatch normalizer) and the true RPM ELBO? Is the reported r=0.9985r=0.9985 tracking a valid bound, or does it also absorb the optimization path?
  • Given that CLIP-posthoc matches or beats sRP-SSM in several settings, what is the minimal regime (noise level, private-variation magnitude, horizon) in which jointly learning dynamics is decisively worth it?
  • Why is the transition term redundant with the filter in the ablation, and does that redundancy persist on the real data? If the filter alone supplies the inductive bias, is the stochastic dynamics model doing independent work?
  • The real-data results come from one mouse. How do you expect the retrieval and decoding advantages to generalize, and can you report multi-animal variability? If not, some discussions are welcome I think.
1 comment

We thank the reviewer for the detailed and constructive review.

Quantify the ELBO approximation: You are correct that the finite minibatch objective is not guaranteed to lower bound the log evidence. The objective differs from the RPM ELBO in two ways. First, it replaces the full normalizer with a finite candidate set (FϕF_\phi vs. F^ϕ\hat{F}_\phi). This error can have either sign and vanishes as the candidate estimate converges to the full normalizer. Second, it applies a Jensen bound to the finite-candidate objective (Eq. 11). This second step is guaranteed to be a lower bound to the finite-candidate objective.

On synthetic data, we estimated the ELBO using a 100,000 candidate reference bank and Monte Carlo posterior expectations as in Figure 2B, but now tracking estimates of both the finite-candidate and Jensen contributions to the minibatch vs. reference ELBO discrepancy. The minibatch objective was below the reference at all 90 checkpoints across five seeds. At the end of training, the total gap was −0.047±0.001-0.047\pm0.001 nats per timestep, with at most 0.01 nats attributable to the Jensen step. We evaluated both objectives at each checkpoint on the same held-out windows with a fixed candidate bank and posterior sampling seed. The reported correlation of 0.99850.9985 therefore measures tracking under a fixed evaluation protocol along the optimization trajectories, while the measured gaps quantify the approximation error. In the revision we will highlight the two sources of RPM ELBO discrepancy throughout training.

Transition term vs. filter ablation Following your suggestion, we performed the same 2×22 \times 2 ablation on real data. We find that joint learning of dynamics via the transition term improves over training with a filter alone in retrieval, latent predictability, and keypoint decoding. With dynamics refit at evaluation, adding the transition loss improved top-1 retrieval by 5.4±0.25.4\pm0.2 percentage points without the training filter and by 1.1±0.51.1 \pm 0.5 points with it (paired across three seeds). The smaller gain with the filter active suggests that the two training signals partly substitute for each other. Please see our response to reviewer WK4L for full details.

When joint learning helps. Across four mice, sRP-SSM and CLIP-posthoc had similar zero-lag retrieval, while sRP-SSM had a longer integrated timescale for future matching by 0.13 - 0.41 s. Within sRP-SSM, training with both the recurrent filter and transition loss increased this timescale by 0.149±0.0310.149 \pm 0.031 s relative to training with neither (paired across three seeds, with dynamics refit at evaluation). These results point to a benefit for matching at future lags. The minimum noise, private variation, or temporal persistence needed for that benefit remains open.

Additional mice: Following your suggestion, we repeated the neural–behavioral experiments in three additional mice, fitting models separately for each animal. Across all four mice, both CLIP-posthoc and sRP-SSM substantially improved cross-modal retrieval over CLIP. The sRP-SSM advantage over CLIP-posthoc varied across mice, with differences ranging from −3.5 to +2.9 percentage points.

The clearest advantage of sRP-SSM was in future cross-modal matching. CLIP-posthoc and sRP-SSM start with a similar zero-lag accuracy, but sRP-SSM is able to match substantially further into the future (0.13-0.41 seconds further across mice as measured by integrated timescale). sRP-SSM also achieved the highest mean keypoint R2R^2 for LFP-to-pose decoding in all four mice. For pose-to-LFP log spectrum decoding, CLIP-posthoc had the highest average R2R^2 in three of four mice.

Additionally, the regional pattern of LFP log spectrum decoding was consistent across mice (Figure 3F). In mice, pose predicted 30-60 Hz power most strongly in the amygdala, the ventral hippocampus, and the nucleus accumbens (R2=0.13−−0.28R^2 = 0.13--0.28). The full region-by-frequency maps correlated with the original mouse's map at r=0.63r=0.63 to 0.810.81.

Reviewer_ux76OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The manuscript proposes a multimodal extension of RPM model family and enable model fit using mini batches rather than computing the ELBO over the full dataset at once.

Critical strengths
  • The authors provide explanations on certain design choices such as the choice of spherical latent geometry and use of RPM model family as the starting point.
  • They show clear advantage of their model formulation on synthetic experiments.
Critical weaknesses
  • Even though the authors provide technical (and theoretical explanations) on certain design choices, it is hard to understand the importance of these design choices for real data applications for non-expert readers.
  • While reading the text, it is hard to grasp why sRP-SSM doesn't rely on reconstruction. I also could not see the ELBO derivation that could help readers understand this point. Could the authors provide that?
  • There exists several multimodal models of neural activity such as PSID (Sani et al., 2021) and MMGPVAE (Gondur et al., 2023), which are not mentioned or included in the comparisons. Further, I could not see any comparisons with supervised models fitted from behavior to LFP signals, which are shown to achieve competitive performance in a recent cross-modal distillation framework (Erturk et al., 2025).
  • The chosen real-world dataset does not necessarily highlight the benefits of the cross-modal retrieval setting which seems to be an important angle of the proposed model. In practice, reconstructing LFP signals from behavioral signals is not as interesting as the other way around, since the neural activity encodes richer sources of variability compared to behavioral signals. Can the authors show the merit of their work in more complex real-world datasets where cross-modal retrieval is more valuable across both directions?
  • The proposed model architecture assumes synchronized modalities, which may not be the case for most of the real world datasets. Have they ever tested their method in such settings? How would their model formulation change to support asynchronous modalities?
Questions for authors

Please refer to weaknesses.

2 comments

We thank the reviewer for their careful review.

sRP-SSM and reconstruction: sRP-SSM doesn't rely on reconstruction in the sense that only encoders or recognition potentials f(m)(z∣x(m))f^{(m)}(z|x^{(m)}) (Eq. 3) are directly parameterized, not decoders or observation likelihoods. The RPM likelihood (Eq. 1) is p0(x)fϕ(z∣x)Fϕ(z)p_0(x)\frac{f_\phi(z|x)}{F_\phi(z)}, where p0(x)p_0(x) is the empirical marginal distribution of observations, and thus parameter-independent.

Applying Jensen’s inequality with the variational posterior q=q(z1:T∣X)q=q(z_{1:T}\mid X) and substituting the RPM joint (Eq. 5) gives log⁡p~θ(X)≥Eq[log⁡p~θ(X,z)q(z∣X)]\log \tilde p_\theta(X) \geq E_q [\log \frac{\tilde p_\theta(X,z)}{q(z \mid X)}] =log⁡p0(X)+∑t,mEqt[log⁡fϕ(m)(zt∣xt(m))Fϕ(m)(zt)]−KL(q∥pψ)= \log p_0(X) + \sum_{t,m} E_{q_t} [\log \frac{f_\phi^{(m)}(z_t\mid x_t^{(m)})}{F_\phi^{(m)}(z_t)}] - \mathrm{KL}(q\Vert p_\psi). The term log⁡p0(X)\log p_0(X) is fixed and drops out of optimization, resulting in Eq. 8.

PSID: We implemented a symmetric version of PSID, finding on synthetic data it gives similar dynamics errors to sRP-SSM and operates near the retrieval/latent recovery frontier. However, on real data, we find it is not competitive on cross-modal retrieval, keypoint decoding, or LFP spectrum decoding. Please refer to our response to WK4L for more details.

MM-GPVAE: MM-GPVAE offers us a chance to evaluate a generative model with partitioned modality-specific and shared subspaces. We evaluated a causal readout adaptation of MM-GPVAE on the synthetic data under the same conditions as VAE-PoE. Hyperparameters were chosen via cross-validation on the training split. With the pruned basis described in the original work, MM-GPVAE improved item retrieval over VAE-PoE at both high SNR (0.767 vs. 0.454) and low SNR (0.230 vs. 0.009), and improved mean single-modality latent R2R^2 from 0.78 to 0.85 and from 0.01 to 0.59, respectively. However, MM-GPVAE is still far from the retrieval/latent recovery frontier in both SNR conditions.

We additionally applied MM-GPVAE to the real mouse data. Requiring MM-GPVAE to reconstruct LFP observations resulted in a collapsed neural latent (KL ≈0\approx 0) and chance retrieval. Including a KL warmup period as in sRP-SSM did not help this optimization failure. A third variant with LFP input but neural decoders required to reconstruct per-channel band powers did not result in a collapsed latent, but again performed cross-modal retrieval at chance levels.

Supervised decoders: Following your suggestion, we also report results from three supervised baseline methods on the pose and LFP decoding tasks (Figure 3D,E): ridge regression, a kk-nearest neighbors regressor, and a matched encoder architecture with an MLP prediction head. All hyperparameters were chosen by two-fold cross-validation within the training split.

We observe poor performance on 3D pose decoding. Averaged over keypoints, ridge, kNN, and the supervised network reach R2R^2 of 0.149, 0.143, and 0.073 respectively, well below sRP-SSM (0.256) and CLIP (0.220). However, on the LFP log spectra decoding task, the supervised network reaches an average R2R^2 of 0.168, above every self-supervised method (sRP-SSM 0.140, CLIP-posthoc 0.141) and the other supervised methods (ridge 0.060, kNN 0.067).

We will include PSID, MM-GPVAE, and supervised decoder results in the updated manuscript.

Neural-behavioral correspondence: Cross-modal retrieval tests whether one modality identifies its paired window in the other. Pose-to-LFP decoding shows which spectral features covary with behavior. We extended the decoding-based rearing analysis to nine additional behaviors (Keypoint MoSeq syllables; Weinreb et al., 2024) across four mice. Decoded and recorded mean LFP spectra around syllable onsets and offsets correlated at r=0.71r=0.71 to r=0.95r=0.95 (+/- 1.5s, 1-100 Hz), in line with the rearing findings. We fit models separately in three additional mice. In all four, sRP-SSM and CLIP-posthoc improved retrieval over CLIP, while sRP-SSM had the highest mean LFP-to-pose R2R^2 and the highest latent predictablity, retaining cross-modal matching information 0.13-0.41 seconds further across four mice as measured by integrated timescale. Evaluation on a richer dataset remains open.

Asynchronous modalities: The proposed method does not rely on simultaneously sampled modalities. The mouse data, for example, consists of LFPs sampled at 1000 Hz and keypoints sampled at 30 Hz, with a camera sync trace used to align the two. A known temporal correspondence is necessary to form positive pairs, however. Missing observations can be naturally handled with the existing modality masking machinery. Irregular observations would additionally require interval-dependent transitions that approximate a continuous-time formulation of the dynamics.

I thank the authors for their explanations, I adjusted my score accordingly.

Reviewer_WK4LOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes sRP-SSM that combines contrastive representation learning with generative state-space modeling to identify shared neural-behavioral dynamics. The method uses spherical latent states and recognition-parameterized observation models to learn shared representations without reconstructing the observations. The authors evaluate the method on synthetic data and mouse neural-behavioral recordings, which shows certain improvement on retrieval and decoding.

Critical strengths
  1. The paper touches an interesting problem of identifying shared dynamics from multimodal time series with substantial modality-specific variability.
  2. The proposed probabilistic formulation is mathematically interesting with the use of spherical latent geometry for tractable evidence fusion and learning.
Critical weaknesses

My concerns are mainly about the insufficient motivation and evidence for the proposed framework.

  1. I did not fully get it why combining contrastive learning with RPM rather than other generative models, like LDS or low-d RNN. sRP-SSM is built on RPM, does RPM provide benefits than other SSM? The author tried to explain the presssure or issues from observation reconstruction in shared neural-behavior modeling, but it is very informal and superficial. I would like to see a strong explanation.
  2. In synthetic experiments, removing the explicity transition objective produces almost identical latent covery, retrieval and dynamics recovery results. And on the real neural-behavioral data sRP-SSM provides 35% retrieval accuracy which is essentially similar to CLIP-postdoc method. These suggests that the joint probabilistic modeling and learning provide weak benefit than just adding a temporal filter.
  3. The paper emphasized the limitations of reconstruction-based generative models. But in the experiments, the authors only compare againsti CLIP variants, and there is not comparison of the proposed method with other dynamical models, for example PSID. Therefore, it is hard to evaluate the so-called reconstruction bias and the benefit of sRP-SSM.
Questions for authors
  1. What is the main advantage of the proposed probabilistic state-space formulation over contrastive representation learning with a simpler temporal filter?
  2. Given the ablation results, can the authors clarify the actual contribution of the explicit transition objective and provide corresponding ablations on real neural-behavioral data?
3 comments

sRP-SSM vs. CLIP-posthoc: Across the original mouse and three additional mice, the zero-lag retrieval difference between sRP-SSM and CLIP-posthoc ranged from −3.5 to +2.9 percentage points. However, future cross-modal matching persisted longer with sRP-SSM, with an integrated timescale 0.13–0.41 s longer across mice. sRP-SSM also achieved the highest mean LFP-to-pose R2R^2 in all four mice. See our reply to riuZ for more details on the additional mice results.

Explicit transition objective: The transition term and the recurrent filter both encourage temporal predictability. The transition term directly rewards consistency with the learned dynamics, while backpropagation through the filter rewards past evidence that remains useful for the current timestep. This overlap explains why removing the transition term had little effect in the synthetic experiments when the training filter was active.

We repeated the ablation on six real-data sessions (mean ±\pm SE across three seeds). With the same evaluation gain and dynamics refit separately for each model, top-1 retrieval was 33.9±0.4%33.9 \pm 0.4\% for full sRP-SSM, 32.8±0.9%32.8 \pm 0.9\% without the transition term, 32.4±0.5%32.4 \pm 0.5\% without the training filter, and 27.1±0.7%27.1 \pm 0.7\% without both. Thus the transition term recovers 5.4±0.25.4 \pm 0.2 percentage points when the training filter is absent and 1.1±0.51.1 \pm 0.5 points when it is active. We also measure the characteristic timescale of the future retrieval results from Figure 3C to see how predictable the learned representations are. CLIP-posthoc has a characteristic timescale of 0.96 s, as measured by integrated timescale, compared with 1.11±0.051.11 \pm 0.05 s for sRP-SSM without the training filter or transition term, 1.20±.021.20\pm .02 and 1.21±.031.21\pm .03 without training filter or transition term, respectively, and 1.26±.021.26\pm .02 s with both. Thus, the transition objective encourages predictable representations beyond what the training filter provides.

PSID: We fit PSID in both directions (LFP to pose and pose to LFP), then align the state estimates on training data by CCA, giving a shared latent space. PSID-CCA uses per-channel log band power on mice. On synthetic data, we found that PSID-CCA achieves similar dynamics errors to sRP-SSM (0.024 vs. 0.021 at high SNR; 0.031 vs. 0.019 at low SNR; both within one SEM) while operating near the retrieval/latent-recovery frontier. However, with a common longer burn-in on mouse data, PSID-CCA performs poorly on three tasks:

  • Cross-modal retrieval: 2.7% top-1 accuracy (chance 0.4%), against 37.2% for sRP-SSM and 34.3% for CLIP-posthoc.
  • 3D pose decoding: average R2R^2 over keypoints of 0.166 vs. 0.256 for sRP-SSM.
  • LFP log spectrum decoding: average R2R^2 over frequencies of 0.099 vs. 0.140 for sRP-SSM and 0.141 for CLIP-posthoc.

RPM, LDS, and low-dimensional RNN. These terms describe different parts of sRP-SSM. The RPM specifies the observation likelihood through modality-specific recognition factors and their aggregate densities (Eq. 1). The latent transition applies an orthogonal linear map on the sphere and uses the resulting direction as the center of a von Mises–Fisher distribution (Eq. 4). The approximate posterior is computed by a low-dimensional recurrent filter (Eq. 6). Our motivation for using an RPM concerns the observation model and the learning signal it gives the shared latent.

We agree that the explanation in lines 291–304 should be more concrete. The minibatch recognition term in Eq. 10 has the form of a one-way InfoNCE objective. In the idealized classification problem, positives are drawn from p(x∣z)p(x\mid z) and negatives independently from the observation marginal p0(x)p_0(x). The optimal score is log⁡[p(x∣z)/p0(x)]\log[p(x\mid z)/p_0(x)], up to an additive function of zz. The RPM parameterizes this likelihood ratio as fϕ(z∣x)/Fϕ(z)f_\phi(z\mid x)/F_\phi(z). To see why this matters, write x=(s,u)x=(s,u), where uu adds no information about zz once ss is known. Then p(u∣s)p(u\mid s) appears in both p(x∣z)=p(s∣z)p(u∣s)p(x\mid z) = p(s\mid z) p(u \mid s) and the marginal p0(x)=p0(s)p(u∣s)p_0(x) = p_0(s) p(u \mid s) and cancels, leaving p(s∣z)/p0(s)p(s\mid z)/p_0(s). The ideal contrastive score depends only on ss, while reconstructing the full observation can reward encoding uu as well.

In a reconstruction-based model whose decoder uses only the shared latent, encoding predictable private details can improve reconstruction and compete for capacity. The RPM objective removes that reconstruction incentive while retaining a probabilistic observation model. In practice, the amount of private information retained can depend on the encoders, finite data, and other details. However, in our synthetic experiments we find that no generative model we tested except for PSID is able to reach the retrieval/latent recovery frontier. PSID is the exception, likely because its identification procedure prioritizes the shared dynamics.

Can the authors be more detailed and specific on what extra experiments and results you have done? I checked the paper and did not find any new results are provided, especially the PSID result you listed above.

Unfortunately, we did not have time to fully update the manuscript with the new results during the author/reviewer discussion period. However, here we have provided additional details for the PSID results described above.

PSID details: After both PSID models are fit, their representations are aligned using CCA with a ridge regularization of 10−410^{-4}. For the synthetic data results, PSID is fit with n1=8 (the true latent dimension), n2=0, and a horizon of 5. Retrieval is computed using unit-normalized state estimates, which we found improved performance.

Synthetic results are scored as in the paper. Mean ±\pm SEM over 3 seeds:

Dynamics error (high SNR)Dynamics error (low SNR)Latent R² (high)Latent R² (low)Retrieval (high)Retrieval (low)
sRP-SSM0.021±0.0090.021 \pm 0.0090.019±0.0080.019 \pm 0.0080.9100.7540.9870.531
PSID-CCA0.024±0.0070.024 \pm 0.0070.031±0.0100.031 \pm 0.0100.9230.7430.9700.557

PSID is linear, so we do not use raw LFP data directly. Instead, for each 300 ms window, we calculate multitaper log band power for each good channel (NW = 1.5, 2 tapers). There were two candidate band sets: 1-4, 4-8, 8-13, 13-30, 30-60, 60-100, 100-150 Hz, or the same bands excluding the 100-150 Hz band. Channels marked bad in any session are excluded, leaving 210, 210, 224, and 217 features for the four mice. For the pose features, we use egocentric keypoints and velocities, either averaged over the window (54 dimensions) or kept per frame (486 dimensions). All features are z-scored with the training split statistics.

We perform model selection on the training set on a grid with 96 hyperparameter sets: 2 band sets, mean or per-frame pose features, horizon in {2,4,8,12}, n1 in {4,8,16}, n2 in {0,16}. There are two training portions from each recording (minutes 0-12 and 18-30), so for each configuration we fit one portion of every session and score on the other, in both directions, and average. The best configuration, chosen by instantaneous retrieval, is refit on the whole training split.

The following validation results are scored using a longer burn-in for warming up the filters, as required by the PSID-CCA fits (44 s for the original mouse, Mouse 1), and so they are not directly comparable to the numbers in the original draft.

Mouse (chance)MethodTop-1 retrieval (%)Pose R2R^2 (mean over keypoints)log spectrum (mean over frequencies)
1 (0.40%)PSID-CCA2.7±0.52.7 \pm 0.50.1660.099
CLIP-posthoc34.3±1.634.3 \pm 1.60.2020.141
sRP-SSM37.2±2.437.2 \pm 2.40.2560.140
2 (0.60%)PSID-CCA5.3±0.85.3 \pm 0.80.2150.151
CLIP-posthoc44.3±3.444.3 \pm 3.40.2610.243
sRP-SSM40.7±5.540.7 \pm 5.50.2980.193
3 (0.48%)PSID-CCA5.9±0.75.9 \pm 0.70.3250.119
CLIP-posthoc52.4±1.552.4 \pm 1.50.3660.185
sRP-SSM52.8±2.452.8 \pm 2.40.4160.186
4 (0.48%)PSID-CCA5.8±0.95.8 \pm 0.90.2320.216
CLIP-posthoc37.5±3.337.5 \pm 3.30.2990.303
sRP-SSM34.3±4.134.3 \pm 4.10.3040.282

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.