acceptodds
Under review as a conference paper at ICLR 2027

Stay on the Manifold: Hierarchical Bayesian Trajectory-Family Transfer for Sparse-Target Reconstruction and Sampling

Abstract

Where should reconstructions and new samples lie when a few observations reveal only fragments of a trajectory? We propose hierarchical Bayesian trajectory principal-manifold sampling (HB-TPM), which learns a family of smooth curves from ordered source tasks and adapts it to sparse targets. For unordered vectors, alternating progression assignments and coefficient updates estimate trajectory structure between and beyond the observed shots under the learned source family. Observation-aware extensions connect the trajectory prior to photon arrivals and image-encoded cardiac cycles. Our analysis characterizes the coefficient directions supplied by the prior and bounds sampling error conditional on curve recovery. Across 30 held-out vector tasks, HB-TPM reduces mean sliced-Wasserstein distance by 21.0–41.8% relative to a source-adapted VAE. Prior-strength and sampling ablations identify coefficient shrinkage as the main supported mechanism. On 100 simulated photon records, adaptation lowers mean relative trajectory RMSE by 50.2% versus target-only fitting and 23.6% versus source-mean prediction. Across three training seeds and five folds on 98 cardiac sequences, HB-TPM with latent-MSE training achieves the best mean reconstruction among the tested configurations. Together, the results demonstrate trajectory-family transfer for sparse-target reconstruction and trajectory-centered sampling when source and target tasks share comparable progression structure.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Hierarchical Bayesian modeling, trajectory-family transfer, principal manifolds, sparse-target reconstruction, few-shot sampling

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Av1gOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes hierarchical Bayesian trajectory principal-manifold sampling (HB-TPM), which learns a family of smooth curves from ordered source tasks and adapts it to targets. The authors performed experiments on two synthetic, vector benchmark and synthetic photon arrivals, and one real dataset, cardiac frame reconstruction. The comparison with the implemented baselines show that the presented method offers the best mean reconstruction.

Critical strengths

The problem, where the new samples lie when a few observations reveal only fragments of a trajectory, the authors investigate is interesting. Current learning methods require a good amount of data to learn the distribution and then sample from it later.

Critical weaknesses

The experiments are weak. The authors performed experiments only on a single realistic dataset. Also, the number of examples generated rom the synthetic datasets are pretty limited. I would expect to see how the method performs with the number of samples changing.

One of the weakest points is that the baselines and overall comparison are in the level of ablations. For example, in Table 3, the authors presented the main result of the realistic dataset but it looks like ablation since the most comparisons are derived from the presented technique.

The other point about the experiments is that the baselines are extremely weak. I would expect to see at least couple of strong deep learning baselines such as VAE or beta-VAE. A comparison with deep learning method while changing the sample size would be very helpful to show the strengths and trade-offs.

The last point is that the overall presentation is poor. There are so much space where the authors can in the main manuscript and bring context from the appendix. The main manuscript is not even complete 9 pages while may details are buried into the supplementary.

Overall, I believe that the manuscript requires a major revision. And, its current experiments are far below the ICLR bar.

Questions for authors

Why not comparing with stronger baselines and using more datasets? The current evaluations are closer to ablations rather than main results.

Reviewer_vxuQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes HB-TPM for recovering trajectories from sparse observations. It learns a prior over curve coefficients from source tasks, adapts the coefficients to a new task, and samples around the fitted curve. For unordered target vectors, it alternates between assigning progression coordinates and updating coefficients. Experiments cover synthetic vector trajectories, simulated photon arrivals, and sparse-frame reconstruction on real cardiac ultrasound sequences.

Critical strengths

The TED experiment provides a real-data application of sparse trajectory reconstruction. It evaluates 98 cardiac cycles using patient-disjoint folds, and the latent-MSE HB-TPM configuration has the best mean reconstruction scores among the tested configurations. This gives the method some empirical support beyond the simulations.

Critical weaknesses
  1. The recovery task may already be explained by a small number of components learned from the source trajectories. The method itself uses a finite-basis representation and fits target coefficients under a learned Gaussian prior (Sections 3.1–3.3). I do not see evidence that a simpler source-trained low-rank model would be insufficient. Estimating unknown target progression coordinates is an additional feature, but the paper needs to show what useful capability this provides beyond fitting sparse observations within a learned population model. Otherwise, it is difficult to judge whether the manifold formulation adds substantial value to the underlying recovery problem.

  2. The vector experiment uses a synthetic eight-dimensional trajectory family, and the photon experiment uses a log-linear diffusion family with two endpoint coefficients (Sections 5.1 and 3.5). Both are controlled settings in which related source and target tasks share a restricted form of variation. They demonstrate recovery within those families, but leave open whether the gains largely reflect that simple shared structure. TED does use real ultrasound, so the experiments are not all synthetic. However, it is one dataset of periodic cardiac cycles and does not by itself resolve whether simpler population models would explain the results. The significance of the broader manifold framing remains unclear to me.

  3. Section 5.1 states that only HB-TPM variants use the source progression coordinates. The comparison with the source-adapted VAE therefore changes both the model and the information available to it. This matters because the central question is whether the proposed framework is needed, rather than whether additional source structure helps recovery. A source-trained low-rank or functional reconstruction method should receive the same available progression information. The paper discusses PACE and Gaussian-process approaches, but Section 6 leaves these comparisons to future work. The reported gains therefore do not establish an advantage over these relevant alternatives.

Questions for authors
  1. What evidence shows that a source-trained low-rank population model, given the same progression information, would be insufficient for these recovery tasks?
  2. Which reported result most clearly demonstrates a benefit from explicit trajectory geometry beyond learning the shared variation of a restricted source family?
  3. How should the vector gains be interpreted given the difference in source-coordinate access, and is there a comparison with a related reconstruction method using the same information?
Reviewer_yqBZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to reconstruct a changing process when only a few observations are available. The main idea is to learn typical trajectory shapes from related, well-observed examples, then adapt that knowledge to a new example with sparse measurements. Rather than simply joining the observed points, the method uses the learned variation across trajectories to estimate plausible missing states and generate additional samples. The proposed method, HB-TPM, uses a hierarchical Bayesian model to balance the shared trajectory pattern against the new observations. For unordered vector observations, it also estimates where each point belongs along the trajectory. Experiments cover synthetic vectors, simulated photon records, and cardiac ultrasound sequences. They show improved reconstruction and sampling in the tested settings, with ablations suggesting that pulling uncertain trajectory coefficients toward the learned prior is the main source of improvement.

Critical strengths
  1. The method reconstructs missing states from very few observations by transferring knowledge from related, well-observed trajectories. - useful in practice.
  2. Its hierarchical Bayesian model learns common trajectory patterns while allowing each target to adapt according to its own observations.
  3. For vector data, it jointly estimates the trajectory and where observations lie, extending beyond reconstruction with known time coordinates. - useful as well
  4. The analysis characterizes which coefficient directions are supplied by the prior and bounds sampling error conditional on trajectory recovery, clarifying the scope of its guarantees. - this analysis strengthens the paper significantly
  5. Experiments span synthetic vectors, simulated photon records, and real cardiac ultrasound. The ablations help identify coefficient shrinkage toward the learned prior as the main supported source of improvement.
Critical weaknesses
  1. The method assumes that source and target trajectories share comparable progression structure. If a target differs substantially, the learned prior could bias reconstruction toward an inappropriate shape. and hence an unfavorable calibration.
  2. Representing each task as a smooth curve is suitable for a single progression variable, but does not directly accommodate branching processes or variation requiring multiple independent coordinates.
  3. Different trajectories and progression assignments may explain the same few points. Alternating updates can select one plausible solution without establishing that it is the correct reconstruction. This comes from a non-bijective nature of this problem, needs to be addresses.
  4. The sampling-error analysis assumes successful curve recovery. It therefore does not fully explain when sparse observations are sufficient to recover the target trajectory, particularly under source–target mismatch.
Questions for authors
  1. How does performance degrade when a target lies outside the learned trajectory family? Can the method detect an unsuitable prior or reduce its influence automatically? Moreover a test of such is useful.
  2. With only a few unordered observations, several trajectories may explain the data. How sensitive are the inferred progression assignments and reconstructions to initialization?
  3. Can you compare HB-TPM with Gaussian-process or neural-process approaches given the same source data and progression information, to isolate the benefit of the proposed transfer mechanism?
  4. Beyond sampling bounds conditional on curve recovery, what evidence establishes when recovery is reliable? Any uncertainty intervals?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.