- Rating
- 3
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The authors proposed an SB method operated in latent space for forecasting of irregularly sampled timeseries in supernovae data. They used SELDON to handle the irregularity and operate most of the probabilistic treatments in the latent space. They used an SB to transport partial LC encodings to full LC encodings, and a separate UQnet for uncertainty quantification. For classification task they poposed a pre-softmax mixing for better calibration.
Critical strengths
- Operating probablistic inference part in latent space is a good idea and SB seems somewhat natural for the partial to full task.
- The authors made good attempt to make sure the inference is calibrated rather than simply being able to generate similar data.
Critical weaknesses
- I think the advantage of doing SB is not well explaind esp. compare to conditional sampling.
- Calibration claims are less supported than point performance.
Questions for authors
- Advantage of SB in place or a conditional diffusion/flow method. The partial to full in SELDON space is done by learning a direct coupling that moves partial representations to full representations.
- There are some computation/learning costs associated with the generality of SB compared to e.g., diffusion which has a known forward process. And the transportation is on the population scale rather than trailered towards any single event. (i.e., the transportation will transport distribution of partial LC's representation to full LC's representation, which is different than trying to get posterior's given any single event).
- Is there a particular advantage so the authors uses SB rather than a conditional sampling method conditioning on partial and sample full? This also closer to a Bayesian inference in the latent space.
- The many to one issue seems can be solved more naturally with conditional sampling than learning a transport?
- Re: Temporal completeness. Will this give the method some look ahead information? Suppose we have a new event can we get it before we observe the full LC?
- Ordering of averaging and decoding in point forecast. The decoding process is non-linear so it does not commute with expectations. So for point forecasting we at least have two options, 1) average the representation, then decode as the authors did or 2) decode all sampled representation, then average. Is the second tried?
- Testing of calibration: the authors claimed that SB's uncertainty is not enough thus a separate UQNet is needed. I am not sure I fully follow if there is an empirical test for this. Please correct me if I missed it.
- Pre-softmax mixing: I assume the hold out data is not used for choosing the 0.85. The authors said it is using a validation data but I am unsure about the relationship between this validation dataset, training set and hold out test set. Also is this tested for claimed better calibration?
- It has nothing to do with the review, but since the authors are likely doing physics and doing linear combination inside softmax, I think the authors will enjoy this connection between what they did and relativity as I did. One sentence summary is that the difference between adding before softmax and after is like adding rapidity vs adding velocity in special relativity: https://betanalpha.github.io/assets/case_studies/general_taylor_models.html#Appendix_A:_Hyperbolic_Geometry
3 comments
3. Empirical Evidence That Bridge Uncertainty Alone Is Insufficient
We apologize if this connection was unclear in the text. The direct empirical test comparing raw I2SB uncertainty against I2SB + UQNet is shown in Table 3 and Table 4:
- Table 3 (UQNet is not used): Even with all bridge variance components active (Config D, which includes completeness-dependent variance and the flux-error floor), raw I2SB trajectories substantially under-cover at the 0.80 nominal level, achieving only 0.305 ± 0.000 all-stage coverage and 0.493 ± 0.007 pre-peak coverage (largely because held-out points in the tails consist of noisy non-detections and the hidden-space bridge does not model observation-level heteroscedastic noise).
- Table 4 (I2SB + UQNet): Passing the hidden-space bridge intervals into our decoupled post-hoc UQNet substantially reduces this under-coverage, lifting empirical coverage at the 80% nominal level from 30.53% to 81.04% (all-stage) and from 49.34% to 74.16% (pre-peak), while leaving the I2SB point forecast unchanged.
4. Ordering of Averaging and Nonlinear Decoding
You are correct that the SELDON decoder is nonlinear, so in general .
Uncertainty quantification. For prediction intervals and coverage metrics, we already decode every sampled trajectory individually,
and compute empirical quantiles directly in flux space, preserving nonlinear decoder effects.
Point forecasting. We additionally compared the two estimators directly:
| Estimator | NRMSE | Held-out RMSE | Latency/object | Throughput | Incremental GPU memory |
|---|---|---|---|---|---|
| D(E[x_full]) | 0.26070 ± 0.00276 | 47.96 ± 1.54 | 7.38 ms | 135.5 objects/s | 12.5 MB |
| E[D(x_full)] | 0.25961 ± 0.00244 | 47.35 ± 1.00 | 15.34 ms | 65.2 objects/s | 3,206.8 MB |
The sample-then-decode estimator gives a small accuracy improvement, whereas decoding the mean representation requires only one decoder pass and is substantially more efficient. We therefore retain for the main point-forecast experiments while using per-sample decoding for all uncertainty estimates. The two estimators achieve nearly identical forecasting accuracy, while roughly halves latency and reduces GPU memory by over 250×.
5. Validation Split Hygiene and Pre-Softmax Mixing ()
Disjoint splits. The dataset is split at the astrophysical-object level into disjoint training (80%), validation (10%), and held-out test (10%) sets, preventing object or partial-view overlap across splits. The fusion weight was selected exclusively on validation pre-peak Macro-F1 and frozen before test evaluation.
Calibration Impact of Pre-Softmax Mixing. Fusion substantially improves calibration relative to the energy-only branch in the pre-peak regime. On the held-out test set, pre-peak ECE decreases from 30.68% at to 3.82% at . Model B alone () achieves a similar ECE of 3.85%. Thus, we interpret fusion as combining complementary signals and substantially reducing the poor calibration of the energy-only classifier, rather than claiming uniform superiority over both single-branch models on every calibration metric.
We will clarify both the validation/test separation and the calibration interpretation in the manuscript.
Thank the authors for the detailed response. It addresses most of my concerns. I would usually increase score once my questions are addressed but this time we only have 4 possible scores and I think general ML contribution-wise the current paper is shy of the top score of 4 so I will keep my 3. To me the paper deserves a raise, just not to 4.
We sincerely thank the reviewer for their thoughtful feedback, constructive questions, and Weak Accept recommendation. We are glad you found latent-space Schrödinger Bridge transport natural for partial-to-full forecasting, and we also appreciated the connection to rapidity addition in special relativity and hyperbolic geometry.
Below, we address each technical question.
1. Advantage of I2SB Over Conditional Diffusion / Flow Matching
You raise an important distinction between population-level transport and event-level conditional sampling p(xfull∣xpartial).
Why I2SB for paired partial-to-full completion. Unlike general Schrödinger Bridge formulations requiring iterative proportional fitting between unpaired marginals, I2SB trains on explicitly paired partial/full representations from the same object using an analytically tractable Brownian-bridge process. Each reverse trajectory starts directly from xpartial, allowing multiple samples for the same partial observation without iterative marginal matching.
Direct comparison with Conditional Flow Matching (CFM). We additionally trained a conditional CFM baseline in the same frozen 256-dimensional SELDON hidden space, with
x0∼N(0,I256)
transported toward xfull while conditioning on xpartial.
These results suggest that CFM is a strong alternative rather than a uniformly weaker baseline, while endpoint-paired I2SB provides favorable behavior in several sparse and low-NFE operating regimes.
2. Temporal Completeness ri Without Look-Ahead Information
Thank you for flagging this issue. Your question prompted us to review our data-loading pipeline, where we found that the original implementation inadvertently read a per-object cached tmax rather than a fixed global value.
We corrected the protocol so that temporal completeness uses no future information at inference. For a new streaming event,
\operatorname{clip} \left( \frac{t_{\mathrm{last},i}-t_{\min}} {t_{\max}-t_{\min}}, 0,1 \right)$$ with $t_{\min}=0$ and a fixed $t_{\max}=221.61\ \text{days}$ defined as the 95th percentile of completed **training-set** light-curve durations. Thus, $r_i$ depends only on the currently observed timestamp $t_{\mathrm{last},i}$. Using this corrected definition, we reran the affected pipeline: - **Three-seed retraining:** I2SB, UQNet, and the temporal classifier were retrained across all three seeds. - **Validation-only fusion selection:** The optimal fusion weight shifted from $\alpha=0.825$ to $\alpha=0.800$. - **Updated results:** Numerical values changed slightly, while the main qualitative trends, UQNet calibration gains, early-stage separation results, and baseline comparisons remained consistent. We will clarify this inference protocol and update the corresponding results in the manuscript.