acceptodds
Withdrawn from ICLR 2027

Calibrating Meta-Learned Feature Maps for Pooled Measurement Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Amortized Bayesian experimental design and context-adaptation methods learn a shared model across many systems, then identify a per-system latent quantity from a handful of observations at deployment. Prior work in this lineage, to our knowledge, assumes each observation directly measures the quantity being modeled. We study a setting where this assumption fails by construction: each observation is a pooled measurement, formed by summing an unknown pairwise interaction law over an experimenter-chosen subset of nodes in a relational system, the same observation model used in quantitative group testing. We show that a shared feature map trained under this pooled model is subject to an identifiability failure that pooling amplifies relative to the pointwise case: training can converge to a reparametrization of the true interaction law that fits the data equally well but is misaligned with the fixed prior used for exact Bayesian inference, producing systematic recovery error even when the feature map's functional form is otherwise correct. We derive the exact condition under which a reparametrization is harmless, and show that the pooled training objective constrains the gauge only weakly: the evidence is invariant to orthogonal transformations and, under sum-pooling, a per-pair signal-to-interference argument suggests diminishing signal against non-orthogonal drift as pool size grows, so finite training leaves ill-conditioned reparametrizations in practice. We mitigate this with a calibration anchor: known ground truth on a small subset of training-system pairs, used to directly constrain the feature map during meta-training rather than the inference-time prior. An orthogonality-penalized variant of the calibration anchor further improves interaction-law recovery accuracy over the baseline at most acquisition budgets tested, on Kuramoto and Winfree coupling.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
amortized Bayesian inference, amortized experimental design, context-adaptation / meta-learning, identifiability, pooled observations, quantitative group testing, interaction law recovery, calibration, representation learning, gauge freedom / reparametrization invariance, prior misspecification, few-shot system identification

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_zgp4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies recovery of pairwise interaction laws from pooled measurements using shared neural feature map and exact Gaussian inference.It analyzes how an invertible transformation of the features changes the latent prior required for correctly specified inference. When retaining a standard Gaussian prior, orthogonal transformations preserve compatibility. The proposed interventions are an auxiliary loss on labeled training-system pairs and an orthogonality penalty computed relative to the true training-system basis. Synthetic Kuramoto- and Winfree-inspired experiments show modest matched-seed recovery improvements, but do not clearly isolate the proposed mechanism from representation misspecification and additional supervision.

Critical strengths
  1. The covariance calculation clearly explains why non-orthogonal feature transformations can miscalibrate a fixed isotropic latent prior, even with exact posterior computation.

  2. Matched-seed experiments show approximately 4% and 7% lower recovery MSE, with improvements on 12/16 and 15/17 seeds and supporting paired tests.

Critical weaknesses
  1. The proposed failure mechanism is not demonstrated. The analysis does not clearly separate an inability to identify the model from difficulty training it. Equation 4 does not establish that pooling makes the harmful transformations harder to learn and changing the latent prior can correct the mismatch under the paper’s assumptions.

  2. The experiments mix several possible causes of improvement. The hidden variation across systems cannot be represented exactly by the shared five-dimensional model. The orthogonality penalty also uses additional ground-truth information. The reported gains could therefore come from improved approximation or extra supervision, controlled comparisons are needed to establish gauge correction as the cause.

Questions for authors
  1. Is the proposed failure an inability to identify the model, or difficulty learning it with limited data and training? How does Equation 4 distinguish these explanations?

  2. Why does Section 3.4 rule out correcting the latent prior, when the covariance transformation in Equation 3 would restore compatibility under the stated assumptions?

  3. Do the improvements persist when the hidden basis-shape parameters are fixed across systems or provided as inputs? This would help separate gauge correction from representation error.

  4. What additional ground-truth information does the orthogonality penalty require beyond the scalar calibration labels and do its gains persist when the comparison methods receive equivalent information?

Reviewer_V9pzOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper considers the setting of meta-learned feature maps and studies why it can predict pooled measurements reasonably well and yet recover the underlying pairwise interactions poorly. The main idea is that there is a misalignment between the learned feature coordinates and the latent prior. The paper they proposes a partial supervision and a orthogonality penalty to improve the recovery.

Critical strengths

The paper is very interesting and well-written, and it seems to address an important problem. However, for someone outside the topic of study in the paper it was a little hard to follow, maybe the paper could benefit from more explanations and intuitions behind the main results and concepts.

Critical weaknesses

The results are very interesting however, it seems that the paper only test with in all experiments with a five-dimensional generative family. It would be also important to evaluate some more on other systems with incorrect latent dimensions and perhaps also include observation-model misspecification.

The linear setting is also interesting and it provides a good baseline to extend beyond the linear setting. However, it would be good it the paper also provide more intuitions on the extensions beyond linear pooled likelihoods and test whether this acquisition improves recovery when the learned posterior is also miscalibrated.

Improved MSE does not consistently track improved conditioning. Can the authors provide more intuitions on that?

There is a double title for the AI USE STATEMENT section in page 9.

Questions for authors

See the weakness section.

Reviewer_D2e6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies pairwise interaction recovery from pooled measurements using a meta-learned feature map and exact Bayesian inference. The authors identify a potential mismatch between learned representations and the fixed Gaussian prior, and propose calibration anchors and orthogonality regularization to improve recovery. Experiments on synthetic Kuramoto and Winfree systems show improvements in prediction accuracy.

Critical strengths
  • The paper addresses a meaningful inverse problem and provides a tractable framework combining meta-learning with exact Bayesian inference.
  • The experiments include multiple baselines and matched-seed comparisons, providing evidence that the proposed regularization improves recovery accuracy in the tested synthetic settings.
Critical weaknesses
  1. The central research problem and identifiability claim are insufficiently established. The authors claim that pooled measurements introduce an identifiability failure absent from pointwise models. However, the derivation in Section 3.2 describes a general mismatch between a reparameterized feature representation and a fixed prior, which can also occur in pointwise linear-Gaussian models. Moreover, Section 3.3 acknowledges that the marginal likelihood is not generally invariant under non-orthogonal transformations. It is therefore unclear whether the identified problem concerns structural non-identifiability, limited observations, or optimization difficulty. The manuscript does not clearly establish what is fundamentally unidentifiable or why the proposed failure is specific to pooled measurements.

  2. The novelty of the proposed solution and its experimental justification need further clarification. Several components of the proposed framework have clear precedents. Choi (2013) studies weighted-graph recovery from additive queries over vertex subsets, closely matching the observation model. Tanaka et al. (2019) combine aggregated observations with shared latent Gaussian processes for cross-dataset inference. Zhang et al. (2020) study learning from aggregate observations and characterize consistency up to equivalence relations. Kirchmeyer et al. (2022), already cited by the authors, establish conditions for identifying new-system parameters using known parameters from representative training systems. Although these studies do not address the exact combination proposed here, the manuscript does not sufficiently establish what new identifiability problem arises specifically from this combination. Furthermore, Appendix F.3 shows that the orthogonality penalty requires access to the true feature basis during training, beyond the labeled pairwise interactions emphasized in the main text. The experiments also do not consistently demonstrate improved gauge conditioning despite improved prediction accuracy, leaving the proposed correction mechanism insufficiently validated. References

[1] Choi, Sung-Soon. "Polynomial time optimal query algorithms for finding graphs with arbitrary real weights." Conference on Learning Theory. PMLR, 2013. [2] Tanaka, Yusuke, et al. "Spatially aggregated Gaussian processes with multivariate areal outputs." Advances in Neural Information Processing Systems 32 (2019). [3] Zhang, Yivan, et al. "Learning from aggregate observations." Advances in Neural Information Processing Systems 33 (2020): 7993-8005. [4] Kirchmeyer, Matthieu, et al. "Generalizing to new physical systems via context-informed dynamics model." International conference on machine learning. PMLR, 2022.

Questions for authors
  1. What is the claimed identifiability failure, and which model quantities are unidentifiable?
  2. How does this failure differ from the same prior–representation mismatch in pointwise models?
  3. Can the orthogonality penalty be implemented without access to the true feature basis?
  4. How do the authors distinguish improved prediction accuracy from correction of the proposed identifiability failure?

Comments on the paper

1 comment
Area_Chair_eDhyOpenReview ↗

Please engage in the discussion

Authors: you have not replied to any raised equation. Please respond to the comments and engage in the discussion.

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

Related papers

Discussion (0)

Sign in to comment.