acceptodds
Withdrawn from ICLR 2027

Risk Aware Compositional Evidence Learning for Heterogeneous Biomedical Observations

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

High dimensional continuous observations provide rich local information for representation learning in biomedical and scientific data. Recent advances in patch based modeling, structured encoding, and feature selection have improved the ability to capture local dependencies. However, existing methods typically aggregate local observations into a single deterministic representation or retain one preferred subset, making it difficult to preserve multiple useful evidence combinations when local information is redundant, heterogeneous, and unequally reliable. Our key idea is to learn an input conditioned, risk aware distribution over multiple evidence compositions rather than searching for a single optimal representation. Based on this idea, we propose Quantized Structured Flow Representation (QSFR). Specifically, QSFR maps continuously varying local observations into reusable discrete evidence units, constructs alternative evidence compositions through flow based sequential sampling, and models composition level utility as a distribution. The estimated utility and utility dispersion jointly reshape probability mass toward informative and reliable evidence states. We evaluate QSFR across multiple heterogeneous cohorts spanning different biological sources and acquisition conditions. The results show consistently favorable representation quality across multiple evaluation criteria. Mechanism analyses reveal improved local stability, expanded compositional support, and lower risk evidence organization.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Compositional Evidence Learning, Structured Representation Learning, Heterogeneous Continuous Observations

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

High dimensional observations are pervasive in many domains such as biomedical or scientific data. Existing emthods usually conduct patch models structured encoding and feature selection all in isolation.  However, existing methods typically aggreagarte observarions The central idea is to learn an input conditioned, risk aware distribution over multiple evidence compositions rather than searching for a single optimal repre- sentation The authors propose Quantized Structured Flow Represen- tation (QSFR) consisting of local patch quantisation (ii)Generative subset sampling (iii)Utility critic

Critical strengths

The problem is well motivated. The challenge of requiring evidence form contextual samples and using encoding and feature selection thorough principled pipeline is relevant and important.

Critical weaknesses

The paper suffers from weak writing and doesnt fexplaining its point clearly. The authors never discuss and often symbols remain either undefined or appear before they are introduced. Prior relevant works i.e Eq (3) is pretty much VQ VAE needs be discussed in structured representation learning. (i)In Section 3.2, the authors explicitly clarify that σϕ\sigma_\phi is "not interpreted as an identifiable estimate of epistemic or aleatoric uncertainty." Yet throughout the title, abstract, and theory, the framework is marketed as discovering "reliable evidence compositions" via "risk-aware" learning. If σϕ\sigma_\phi does not capture identifiable epistemic or aleatoric uncertainty, the formal justification for treating it as a proxy for physical "reliability" or clinical "risk" is ungrounded.

(ii)Because the critic is supervised using an observed empirical utility margin per sample, it is unclear why the spread σϕ\sigma_\phi of its output distribution should correlate with sample or feature unreliability. The authors provide no empirical validation showing that high-σϕ\sigma_\phi subsets correlate with higher held-out error, higher out-of-distribution frequency, or label noise.

(iii)Theorem 3.1 asserts a supremum bound over all subsets S∈SxS \in \mathcal{S}_x while relying on an average quantization distortion δq\delta_q. Furthermore, assuming a global Lipschitz constant LFL_F for the end-to-end composite map (which incorporates a deep contextual encoder RωR_\omega) is unrealistic for deep neural networks without explicit spectral normalization.

Questions for authors

(i) Does σϕ\sigma_\phi correlate with prediction error or out-of-distribution spectra? What are the quantitative results across Table 1 when α=0\alpha = 0 (pure mean utility)?

(ii)How are GxG_x and BxB_x formally partitioned in real Raman data? Can you provide empirical validation demonstrating that Δσ>0\Delta_\sigma > 0 and Γα>0\Gamma_\alpha > 0 hold in practice?

(iii)What is the active codebook perplexity and utilization percentage across the cohorts? How does one manage codebook collapse in the absence of an autoencoding reconstruction objective?

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

This article explores how to select and combine useful information from complex biomedical signals. The author believes that a signal may have multiple sets of local features that can support prediction, but the reliability of these combinations varies, so only one set should be retained. To this end, the paper proposes QSFR: first convert continuous signal segments into reusable discrete representations, then use GFlowNet to gradually generate different feature combinations, and finally, based on the prediction performance and fluctuation degree of each feature group, tend to choose the combination with better performance and more stable performance. The paper conducted experiments on eight Raman spectroscopy data queues and supplemented with component ablation, domain method comparison, and cross queue testing. The overall results show some potential, but it is still unclear whether the improvement in performance is due to the model's ability to generate multiple combinations and control risks, or mainly comes from discrete coding, stronger prediction modules, or additional training supervision.

Critical strengths

The paper proposes QSFR: first, convert continuous signal segments into reusable discrete representations, then use GFlowNet to gradually generate different feature combinations, and finally, based on the prediction performance and fluctuation degree of each feature group, tend to choose the combination with better performance and more stable performance. The paper conducted experiments on eight Raman spectroscopy data queues and supplemented with component ablation, domain method comparison, and cross queue testing.

Critical weaknesses

The true innovation of the method needs to be clarified. The several main techniques used in QSFR, including vector quantization, stepwise selection of feature subsets, GFlowNet training, and adjusting selection preferences based on predicted risk, have a foundation in related research. The possible contribution of this article is to organize them into a unified biomedical signal processing process, but it is currently not fully explained which step is new.

The paper emphasizes "retaining multiple combinations", but it is still unclear how to use these combinations during testing. The implementation instructions in the appendix mention that during testing, it seems that a greedy strategy was used to select a set of features; And the algorithm description also uses a sampling process. If the final prediction relies only on a fixed set of features, then the classification results can only indicate that this training method helps find good features, and cannot directly prove that "retaining multiple combinations" itself has advantages.

  1. Whether risk punishment is truly effective still lacks targeted ablation experiments. The paper estimates the returns and volatility of different feature combinations through a utility prediction module, and reduces rewards for combinations with greater volatility. But the current experiment of "removing UCRC" has changed multiple factors at once: it has not only eliminated volatility penalties, but also changed profit prediction and auxiliary training. Therefore, even if performance declines, it cannot be determined whether risk control has played a role.

  2. The amount of data may seem large, but the number of independent patients or specimens needs to be explained. The paper used 55458 spectra, but multiple spectra may come from the same patient or specimen and cannot be directly treated as independent samples. For example, the Myeloma SERS queue only has 26 urine samples. It is reasonable to divide the training and testing sets according to patients or specimens in the paper, but there may still be significant fluctuations in the statistical results of some datasets. It is recommended to list the number of independent patients or specimens, the number and distribution of categories, as well as the number of independent units included in the test set for each dataset, and provide the uncertainty of the results in terms of patients or specimens.

  3. The representation and reproduction details of feature combinations still need to be supplemented. The paper aims to treat the selected local features as a "set", but the implementation also uses GRU that is sensitive to input order.

Questions for authors
  1. Has a new risk measurement method, a new generative training approach been designed, or has an effective combination of these existing technologies been proposed? At the same time, it is necessary to strengthen the comparison with the closest work such as risk sensitive GFlowNet, and avoid describing methods that are larger than the actual scope of innovation.

  2. It is necessary to unify the algorithm and implementation instructions, and compare them under the same computational budget: selecting the optimal group each time, randomly sampling one group, taking the average of predictions after sampling multiple groups, and ordinary deterministic feature selection. At the same time, it is best to directly check whether the actual combinations collected by the model are more focused on high reward candidates, rather than just looking at the final accuracy.

  3. Can we keep the rest of the structure and training process unchanged, only turn off the penalty for utility fluctuations, and compare it with the complete model?

  4. It is recommended to list the number of independent patients or specimens, the number and distribution of categories, and how many independent units are included in the test set for each dataset, and provide the uncertainty of the results in terms of patients or specimens

Before selecting local features, the model has already allowed different segments to exchange information. Therefore, the final selected features may contain information from unselected segments. If the paper interprets the results as' predicting based solely on these local pieces of evidence ', it needs to explain this limitation or test the impact of unselected information through experiments

Reviewer_A9hMOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes QSFR for classifying high-dimensional continuous biomedical signals, mainly Raman/SERS spectra. The signal is split into patches that are vector-quantized against a learned codebook (QSU) and contextualized into "evidence atoms." A GFlowNet trained with trajectory balance (CSFG) builds subsets of atoms, and a classifier head predicts from the mean-pooled subset. A distributional critic (UCRC), a categorical distribution over 51 utility atoms, is trained to predict the head's own label margin for each subset; its mean and std define the GFlowNet reward exp{β(µ − ασ)}. Two theorems bound the TV distance between the learned and an oracle subset distribution and show that high-σ subsets get exponentially less mass. Experiments on eight Raman cohorts report the best performance on most cohorts, plus ablations, a transfer experiment, and qualitative "before/after" utility–dispersion plots on Raman, ECG and EEG.

Critical strengths
  1. Real, diverse clinical Raman data with patient-level splits. Eight cohorts spanning serum, urine, platelets, bacterial isolates and airborne cells, with patient/specimen-level splitting and five seeds, is a useful and uncommon test bed for spectral representation learning. This is the main reason the empirical section is worth reading.
  2. The stop-gradient between critic and flow is a sensible design choice. Training UCRC only on its supervised target and feeding the GFlowNet a detached reward avoids the obvious failure where the critic inflates rewards to reduce the TB loss. It is a small but correct engineering decision.
Critical weaknesses
  1. I have to say the core mechanism is not what drives predictions, and the theory is close to tautological. The paper's central claim is learning a distribution Π(S|x) over multiple evidence compositions instead of a single representation. But inference is "greedy until T_max": one subset, one prediction. The distribution is never used at test time, so the benefit of "preserving multiple compositions" is not what the accuracy numbers measure.

  2. The experimental evidence is not reliable enough to support the claims.

(a) Myeloma has 26 specimens, so a patient-level 8:1:1 split leaves 2–3 test specimens; Ovarian (55 patients) and Hepatic (65 patients) leave roughly 5–7 test patients each. With a single split, the ± values reflect seed noise only, not split variance, so the large gaps in Tables 1/7 are not statistically meaningful.

(b) On Platelet SERS, every baseline accuracy is a multiple of 100/120 (74.17, 75.83, 86.67, 90.00, …), consistent with 24 test spectra × 5 seeds. QSFR's 94.31 ± 0.92 is not a feasible value on that grid, and a 0.92 std is impossible when one sample changes accuracy by 4.17 points. Recall 85.00 ± 0.20 has the same problem. This suggests QSFR and the baselines were not evaluated on the same test set.

(c) On Myeloma, the "Base" GFlowNet scores 64.2%, while a plain MSGNet gets 93.9%. Removing only the σ term (w/o UCRC) costs about 20 points even though inference is greedy and α = 0.5. Without exact ablation definitions, this looks more like broken ablation variants than evidence for the components.

(d) The mechanism probes (Fig. 3C, ECG/EEG) compare "before" vs "after" training with no definition of "before" and no predictive numbers, so they show only that a trained model is more confident than an untrained one.

Questions for authors
  1. Inference uses a single greedy subset. Have you evaluated sampling multiple subsets and aggregating? If accuracy is the same, what evidence remains that learning a distribution matters?
  2. Platelet SERS: how many test spectra per seed, and were QSFR and all baselines evaluated on the identical split? Please reconcile 94.31 ± 0.92 accuracy and 85.00 ± 0.20 recall with the granularity implied by the test-set size.
  3. How many test patients/specimens are in each cohort's test split? Can you report patient-level cross-validation (or repeated splits) instead of a single split with five seeds?
  4. Please define precisely what "Base," "w/o UCRC," "w/o CSFG" and "w/o QSU" contain. Why does removing only the σ penalty (α = 0) drop Myeloma accuracy from 94.5 to 74.2 when inference is greedy?
  5. What are M, T_max and |S_x| in practice, and what is the empirical range of µ − ασ? Is the learned distribution actually concentrated, or near-uniform, at β = 1?
  6. What exactly is the "Before" model in Fig. 3C?

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

Related papers

Discussion (0)

Sign in to comment.