Stay on the Manifold: Hierarchical Bayesian Trajectory-Family Transfer for Sparse-Target Reconstruction and Sampling
Organizations: Department of Biostatistics University of California, Los Angeles Los Angeles, CA 90024 · Department of Computer Science Zhejiang University Hangzhou, China · Zhejiang Lab Hangzhou, China
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.
Figures & tables
| Method | |||
|---|---|---|---|
| Gaussian shrinkage | |||
| Pairwise interpolation | |||
| Naive TPM | |||
| Target-trained VAE | |||
| Source-adapted VAE | |||
| HB-TPM (normal) |
| Method | Mean error | 95% CI | Successful fits | Coverage |
|---|---|---|---|---|
| Target-only event QMLE | 0.4378 | [0.3875, 0.4927] | 91/100 | 82% |
| Source mean (post-hoc) | 0.2855 | [0.2568, 0.3163] | n/a | — |
| OA-HB-TPM | 0.2180 | [0.1928, 0.2450] | 100/100 | 91% |
| Method | Dice | HD95 (px) | Area RMSE | 95% coverage |
|---|---|---|---|---|
| Weak-prior adaptation | 0.7454 | 10.0020 | 0.6926 | 43.6% |
| Fixed source coefficients | 0.7472 | 9.8771 | 0.6622 | — |
| HB-TPM (latent-MSE training) | 0.7519 | 9.4843 | 0.6357 | 63.6% |
| OA-HB-TPM (original) | 0.7461 | 9.9467 | 0.6728 | 39.3% |
| OA-HB-TPM (isotropic prior) | 0.7461 | 9.9472 | 0.6727 | 39.1% |
| OA-HB-TPM (no phase alignment) | 0.7462 | 9.9459 | 0.6726 | 39.3% |
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset / setting | Size and split | Estimator-visible input | Prediction and comparison |
|---|---|---|---|
| Synthetic 8-D trajectories | 48 sources 96 ordered points. 30 held-out targets, . | Source coordinates and vectors for HB-TPM. Target vectors have no coordinates. | Sample distribution and curve recovery. Gaussian, pairwise, naive TPM, VAEs, and HB ablations. |
| Synthetic photon arrivals | 40 source and 100 target records. Prior fitted from 39 successful sources. | Event times, duration, and calibrated observation parameters. Clock time is observed. | Diffusion trajectory. Target-only event QMLE, source mean, and OA-HB-TPM. |
| TED ultrasound ( Painchaud et al., 2022 ) | 98 patients, five patient-disjoint test folds, three training seeds. 24 frames per cycle. . | Sparse target images and times. Source training uses masks. Hidden images and target masks are excluded. | Hidden-frame masks and area curves. Six original configurations plus OA observation-noise scale 4.0. |
| Synthetic Gaussian control | 30 repeats per budget, . No HB-TPM run. | Sparse target vectors. Naive TPM also uses an oracle noise scale. | Distribution recovery by Gaussian shrinkage versus naive TPM. Off-trajectory diagnostic only. |
| Synthetic spiral pilot | Three classes. per class. 2,100 test points, one seed per cell. | Target-only labeled shots. No source trajectory prior. | Classification by seven recorded prototype methods. Reproduced, but not compute-matched. |
| MNIST / face image studies ( LeCun et al., 1998 ; Liu et al., 2015 ) | Earlier figure and summary records. Exact splits, identities, and checkpoints are incomplete. | Encoded static images under the earlier target-only workflow. | Qualitative geometry and interpolation, reported smoothness and classification. Not tests of hierarchical transfer. |
| Method | |||
|---|---|---|---|
| Gaussian shrinkage | |||
| Pairwise interpolation | |||
| Naive TPM | |||
| Target-trained VAE | |||
| Source-adapted VAE | |||
| HB-TPM (normal) |
| Method | |||
|---|---|---|---|
| Gaussian shrinkage | |||
| Pairwise interpolation | |||
| Naive TPM | |||
| Target-trained VAE | |||
| Source-adapted VAE | |||
| HB-TPM (normal) |
| Method | |||
|---|---|---|---|
| Gaussian shrinkage | |||
| Pairwise interpolation | |||
| Naive TPM | |||
| Target-trained VAE | |||
| Source-adapted VAE | |||
| HB-TPM (normal) |
| Method | |||
|---|---|---|---|
| Gaussian shrinkage | — | — | — |
| Pairwise interpolation | — | — | — |
| Naive TPM | |||
| Target-trained VAE | — | — | — |
| Source-adapted VAE | — | — | — |
| HB-TPM (normal) |
| Gaussian shrinkage | Naive TPM | Gaussian wins | |
|---|---|---|---|
| 3 | 26/30 | ||
| 5 | 29/30 | ||
| 10 | 27/30 |
| Quantity | Generating family | Learned prior |
|---|---|---|
| ( ) | 45.00 | 63.50 |
| ( ) | 95.00 | 88.67 |
| Log-start variance | 0.0324 | 0.1532 |
| Log-end variance | 0.0484 | 0.0992 |
| Endpoint correlation | 0.650 | 0.041 |
| Method | ||
|---|---|---|
| No augmentation | 60.29 | 56.86 |
| Gaussian augmentation | 63.10 | 57.43 |
| Pairwise interpolation | 58.90 | 56.52 |
| Random-order curve | 49.10 | 56.90 |
| PCA line | 56.71 | 57.76 |
| Offline TPM prototype | 56.14 | 58.57 |
| Digit | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 |
|---|---|---|---|---|---|---|---|---|---|---|
| Raw input | 11 | 9 | 12 | 12 | 12 | 13 | 11 | 10 | 14 | 11 |
| AE feature | 10 | 8 | 10 | 11 | 10 | 11 | 9 | 9 | 11 | 10 |
| TPM feature | 4 | 3 | 5 | 5 | 5 | 5 | 4 | 4 | 5 | 4 |
| Classifier training data | Evaluation data | Reported accuracy |
|---|---|---|
| Real spiral data | Generated samples | 97.89% |
| Generated samples | Real spiral data | 89.4% |