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