acceptodds
Under review as a conference paper at ICLR 2027

Learning Across Feature Dimensions via Dimension-Conditioned Latent Schrödinger Transport

Abstract

We study supervised learning when observations of the same type are represented by different numbers of features. The training data are grouped by feature dimension, and cross-dimensional sample identities are not used. The objective is a single predictor that operates on all training dimensions and transfers to dimensions not used for fitting. We introduce a dimension-conditioned encoder followed by a conditional latent transport to a common reference distribution. The transport is formulated as a soft-constrained multi-marginal Schrödinger bridge and approximated by conditional flow matching. A single task head is then trained on the transported representations. We give a uniform score-error decomposition, a dimension-coverage bound under an assumed per-dimension estimation rate, and a penalty-convergence result for the ideal soft bridge. Synthetic experiments examine shared prediction, dimension coverage, and finite-network penalty behavior. In an independent PBMC3k benchmark retraining, the final configuration attains mean accuracy , macro-F1 , and macro-AUROC across six feature dimensions. Dimension-specific classifiers remain stronger in accuracy and F1, which identifies the present cost of using one model across dimensions.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Cross-dimensional learning, domain generalization, latent transport, Schrödinger bridge, single-cell classification

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3bNZOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses supervised learning across datasets with varying feature dimensions without requiring cross-dimensional sample correspondences. The authors propose a dimension-conditioned encoder followed by latent transport to a common reference distribution, enabling a shared predictor across both training and unseen dimensions. The transport is formulated using a soft-constrained multi-marginal Schrödinger bridge and approximated through conditional flow matching with Sinkhorn-based penalties. The paper provides theoretical results on prediction-score approximation, dimension coverage, and convergence of the idealized bridge formulation. Experiments on synthetic regression and PBMC3k classification demonstrate the feasibility of shared prediction and show that staged optimization performs better than the evaluated joint-alignment alternatives, although dimension-specific models remain stronger.

Critical strengths
  1. The paper addresses an interesting problem of learning across variable input dimensions without paired observations. The proposed framework combines dimension-conditioned representations, latent transport, and a shared task head, providing a coherent approach that can also be evaluated at unseen feature dimensions.

  2. The PBMC3k experiments demonstrate that stronger distributional alignment does not necessarily improve classification performance and that staged optimization provides better stability and predictive performance than the evaluated joint-alignment variants. The authors also clearly distinguish the scope of their theoretical guarantees from the practical neural implementation.

Critical weaknesses
  1. Although several training configurations are compared, the experiments do not sufficiently isolate the contribution of conditional latent transport. A direct comparison with a dimension-conditioned encoder and shared classifier without transport, under comparable model capacity and training budgets, would better establish the benefit of the proposed approach. The performance gap relative to dimension-specific predictors, particularly in macro-F1, also indicates room for improvement.

  2. The theoretical results rely on assumptions that are not established for the proposed learning algorithm. In particular, Theorem 5.5 assumes smooth variation of population risk across feature dimensions, while Theorem 5.7 concerns the idealized Schrödinger bridge rather than its flow-matching approximation. Although these limitations are acknowledged by the authors, a stronger connection between the theoretical analysis and empirical generalization would improve the contribution.

Questions for authors
  1. Can the authors provide an ablation comparing the full method against a dimension-conditioned encoder with a shared task head but without conditional transport, using comparable training budgets?

  2. How closely does the flow-matching objective in Equations (10)–(13) approximate the ideal Schrödinger bridge in Equation (9)? Is there evidence that this formulation offers advantages over standard conditional flow matching?

  3. How sensitive is performance to the distance between training and held-out feature dimensions? Can the authors provide empirical support for the dimension-regularity assumption used in Theorem 5.5?

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

This paper studies supervised prediction across different input feature dimensions, including inference dimensions don't align with training. A encoder mapping inputs into a common latent space is obtained by a autoencoder training process, followed by a conditional flow that transports their distributions toward a distribution whose dimension was highest. The flow is trained with flow matching and marginal constraints on intermediate distribution, and a shared prediction head is then trained on the transported representations.

Critical strengths

The problem is meaningful, and the authors provide a workable pipeline with some experimental support. Some individual designs also make sense. Some experimental support and theoretical analysis are provided. Several individual design choices are understandable.

Critical weaknesses

My main concern is that the paper’s central contribution is difficult to identify. The method combines several established components, but the particular insight and gains from this combination could be clearer.

The intermediate-marginal constraint seems especially important to clarify. It is unclear why input dimensions should determine positions along the transport path, or how this constraint improves prediction on unseen dimensions. Figure 1c examines distribution matching rather than predictive benefits. Table 2 supports the overall training scheme, but several settings change together, making the contribution of this constraint hard to assess.

Questions for authors

Could the authors explain the motivation for the intermediate-marginal constraint and its benefit for prediction on unseen dimensions?

What roles do QQ and the coupling encoder–decoder architecture play? Why are they suitable choices for this task?

What motivates the reference distribution and the noisy interpolation path?

Could the authors provide comparisons with and without transport, and with and without the intermediate constraint, under otherwise comparable settings? Results on unseen dimensions would help clarify the predictive value of these designs.

The problem is meaningful, and the authors provide a workable pipeline with some experimental support. Some individual designs also make sense. However, after reading, I find the main contribution hard to identify. There are several familiar blocks, like autoencoders, alignment, distance preservation, optimal transport, and flow matching. I am not clear about what combining them adds to this problem. And the motivations behind several design choices are not clearly explained. Hence, I tend to give a score 2.

1 comment
Reviewer_kQJwOpenReview ↗

Final Assessment

Thanks for authors' detailed response. I was initially between 1 and 2. After reading the authors' response, I settled on 2. The authors explain why nested feature sets have a natural order and show that the combined method helps in some settings. My main concern is that the contribution is largely an engineering effort to combine existing methods. I came away with little new understanding of when or why this combination helps with learning across feature dimensions.

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

This paper studies how to train one predictor for datasets where different instances of a given type may have different associated feature dimensions. Rather than training a separate model for each feature dimension, the authors propose a "dimension-conditioned encoder" that maps each input into a latent space, followed by a conditional transport model that aligns representations to a common reference distribution. Lastly, a single classification head predicts labels from the transported representations. The paper provides three theoretical results: a decomposition of prediction-score error into task-head and representation errors; an expected risk bound for unseen feature dimensions; and a convergence result for the ideal soft-constrained bridge. Experiments are presented for two scenarios: a synthetic regression task, and PBMC3k single-cell classification across six gene-feature dimensions. Results show that the shared predictor does transfer to unseen held out dimensions, but it significantly underperforms dimension-specific classifiers.

Critical strengths

Formulation of encoder->transport->regressor/classifier approach for the class of problems where feature dimension varies for observations of a given type

Critical weaknesses
  • The biggest weakness is lacking empirical support for the proposed approach. Only two tasks are studied and one of them is synthetic with maximum 16 dim inputs. Results show that in both tasks a dimension specific classifier/regressor significantly outperforms the shared one. Given that, how is the proposed shared model justified?
  • Paper is difficult to read and needs details. For instance:
    • What is the equation for Lgeo\mathcal{L}_{geo} mentioned on page 3?
    • Details of models used for encoder (and decoder that supports encoder training), transport model, and classifier/regressor? What are the model architectures?
  • No comparisons with any other approaches is carried out. For example, transformers are quite good at handling input dimension variability, would they be a plausible alternative model family to consider and compare with?
Questions for authors

please see weaknesses above

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.