acceptodds
Under review as a conference paper at ICLR 2027

Variational Manifold Alignment for Noisy Multilabel Prediction Correction

Abstract

Label noise is ubiquitous in machine learning applications and usually leads to a geometric displacement in a trained classifier’s latent space. We propose LSNPC (Latent-Shift Noisy Prediction Correction), a post-hoc correction method that refines the multilabel predictions of a classifier trained on noisy labels, without retraining it. LSNPC is a variational autoencoder that aligns clean and noisy label distributions on a shared latent manifold by decoding both through the same decoder; this alignment is the primary corrective mechanism, which we design so that decoding a noisy prediction recovers most of the lost signal. A heavy-tailed Student-t transition models the noise-induced displacement, and a Gaussian shift from the noisy toward the clean representation then closes the residual displacement, yielding a smaller secondary refinement. We prove that the learned correction posterior approximates the true posterior over clean representations, that the Student-t shift yields a Kullback-Leibler divergence bound linear in the label discrepancy (a Gaussian alternative adds a quadratic term), and that the label-covariance reconstruction error is controlled by the latent displacement and label sparsity. Experiments on four synthetic-noise benchmarks, with real-world evaluation in the appendix, show that LSNPC improves prediction performance in most settings, with the largest gains observed under severe corruption.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Noisy multilabel learning, Post-hoc noisy prediction correction, Variational AutoEncoder

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rF12OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LSNPC, a post-hoc method for correcting noisy multilabel predictions. It uses a VAE with a shared decoder, Student-t latent variables, and a Gaussian correction step. It supports both unsupervised and semi-supervised training. The paper provides theoretical analysis and tests on four synthetic-noise benchmarks and DeepFashion. Results show frequent F1 gains, especially under severe noise. Improvements in mAP are less consistent.

Critical strengths
  1. On VOC07 with 50% symmetric noise, the ResNet-50 MLP improves from 58.7 to 67.4 CF1 (Tab. 1). This is a meaningful gain for an existing classifier.

  2. The tests cover several classifiers, backbones, and noise settings. On DeepFashion, mAP improves for all six baselines. This provides evidence beyond synthetic corruption.

Critical weaknesses
  1. They derived an affine bound with a generally positive intercept: D_KL \le C_1+C_2\Delta. This alone does not establish the stated vanishing linear rate. The assumed domain gives 0 \le \Delta \le 1 ---> \Delta^2 \le \Delta. Thus, the Gaussian bound also admits a linear upper bound. The different displayed forms do not establish a Student-t advantage. The tail-mass bound in Appendix D does not follow from Definition 3.5 for multilabel probabilities, which need not sum to 1.

  2. Direct decoding captures most of the gains as shown in Tab. 7. The separate-decoder variant leaves the clean decoder untrained, so its failure is expected. A matched single-latent autoencoder and simple score calibration are missing. These controls matter because F1 often improves while mAP stays flat or declines.

Questions for authors
  1. Can you correct the Student-t/Gaussian comparison and show whether an advantage remains when constants and intercepts are included? How should Definition 3.5 change to justify the tail-mass bound?

  2. What exact loss is implemented? The KL decomposition below Eq. (1) appears to omit the dependence between the two latent variables. Does Theorem 3.1 apply to the actual reweighted training loss?

  3. How does LSNPC compare with a single-latent autoencoder and per-class score calibration under matched features, training budgets, and access to clean labels?

Reviewer_8KfeOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper focuses on correcting the predictions of a multilabel classifier trained on noisy labels, without retraining it. The authors propose LSNPC, a VAE that treats label noise as a displacement in latent space: it encodes the image and the noisy prediction into a heavy-tailed Student-t latent, shifts it toward a clean latent, and decodes both through a shared decoder, which is what does most of the correcting. Experiments on four benchmarks show CF1 and OF1 gains in most settings, largest at 50% noise, with a semi-supervised variant improving further.

Critical strengths
  • The core idea is appealing: forcing clean and noisy labels onto a shared latent space through one decoder is a natural assumption that holds broadly and is worth investigating, and pairing it with a Student-t transition is a sensible way to cover widely varying noise levels in a single model.
  • The experiments are broad and the gains are consistent, holding across datasets, base classifiers, and both noise types.
Critical weaknesses
  • There are no experiments showing that label noise actually produces the heavy-tailed latent displacement the method assumes, even though the corruption is injected under controlled conditions and the clean labels are retained.
  • The paper reports only aggregate F1 and mAP, so it is impossible to tell how many flipped labels the correction recovers against how many correct ones it breaks, or which classes it handles worst.
  • The noise is injected independently per label, so it carries none of the label correlations the paper mentions as what makes the multilabel setting hard.
Questions for authors

Since the corruption is injected under controlled conditions, how many flipped labels the correction actually recovers against how many correct ones it breaks?

Reviewer_NbViOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes LSNPC (Latent-Shift Noisy Prediction Correction), a post-hoc framework designed to correct the predictions of a pre-trained multilabel classifier trained on noisy labels without requiring retraining.

Critical strengths

Retraining large models (like modern Vision Transformers or multi-modal backbones) on cleaned datasets is computationally prohibitive. A post-hoc corrector that trains quickly on classifier outputs without touching model weights is of real practical value.

Critical weaknesses

The method is effectively just an autoencoding denoiser through a bottleneck. The elaborate latent transition math is somewhat disconnected from practical performance. able 7 admits that direct-z^\hat{z} decoding recovers 84% to 98% (and up to 100% in semi-supervised) of the entire gain! meaning the complex latent-space shift z(hat) to z contributes almost nothing in practice.

In multilabel classification, mAP (mean Average Precision) is arguably the most important metric because users rank candidate labels or items. - the paper claiming state-of-the-art correction shouldn't degrade ranking quality without stronger mitigation. The authors admit the Bernoulli VAE "sharpens" probabilities toward 0 and 1, helping threshold metrics p > 0.5 but scrambling ranking order.

Most of the dramatic improvements occur at 40% and 50% synthetic noise. At 30% noise (or on cleaner, realistic datasets like DeepFashion), the gains are very modest (often < 1 F1 point) or negative. At 50% multilabel noise, the dataset is essentially coin flips, making it an unrealistic benchmark for deployed production systems.

Questions for authors

how much of the theoretical machinery surrounding the transition mapping and Gaussian recentering is practically necessary versus an autoencoder with a shared decoder? (Table 7 demonstrates that 84–98% of the performance gain is achieved simply by decoding directly from z(hat) without the latent shift.

Since mAP is the primary ranking metric in multilabel applications, the degradation of mAP due to confidence sharpening under moderate symmetric noise is concerning. Did the authors explore entropy regularization or ranking-based losses (e.g., listwise margin ranking) in the ELBO to prevent within-class rank scrambling?

Why does the method exhibit sub-point regressions on well-calibrated baseline models (like ASL) at 30% noise? Does the confidence sharpening interfere with classifiers that already feature well-calibrated posterior probabilities?

Can the authors provide runtime and memory profiling showing the post-hoc inference latency introduced by sampling latents during a production forward pass compared to the base classifier alone?"

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.