acceptodds
Under review as a conference paper at ICLR 2027

Label Provenance Learning: Modeling Where Noisy Labels Come From

Abstract

Noisy-label learning commonly represents supervision as input–label pairs, discarding the source of each label. Yet large datasets pool labels from different websites, collection protocols, or human annotators, whose errors need not follow the same pattern. We study this setting as label provenance learning (LPL): the learner observes , where is a potentially corrupted label, identifies its source, the clean label is latent, and no source is assumed trusted. Provenance can act through two channels, a routing channel describing how examples reach sources and a labeling channel describing how sources produce labels; each may or may not depend on the provenance, defining four nested prediction structures. We derive provenance-conditioned variational learning (PCVL), which fits any of these structures under a single objective, and a permutation diagnostic that selects among them before training, using no clean labels. On CIFAR-10N the diagnostic detects neither dependency, and the provenance-blind structure has the highest mean accuracy among the four; on dopanim it detects a source-dependent labeling channel, and activating only that channel improves balanced accuracy over both provenance-blind PCVL and DivideMix. The diagnostic's decisions agree with post-hoc oracle audits in every real-data setting. A controlled CIFAR-10 cross that varies the two mechanisms separately shows the labeling channel tracking its generating mechanism, with weaker evidence for routing; a matched stress test shows that shuffling provenance removes the gain. Provenance is therefore worth modeling when its conditional role is supported, rather than activated by default.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Learning with Noisy Labels, Weakly Supervised Learning, Deep Generative Models, Label Provenance

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_crEXOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies noisy-label learning with recorded label provenance. Each training example contains an input, one observed label, and a source identifier; the clean class is latent and no source is trusted. The authors distinguish a routing channel, describing source assignment, from a labeling channel, describing label production. Nested restrictions yield four PCVL structures. A provenance-blind warm-up representation supports stratified permutation diagnostics that select a structure before the variational branch is activated. Training combines the generative objective with DivideMix-based semantic supervision, while deployment uses only the input. Experiments include a controlled 2×2 data–model cross, true-versus-shuffled provenance, CIFAR-10N, and dopanim. The main finding is conditional utility: CIFAR-10N favors the provenance-blind form, whereas dopanim benefits from source-dependent labeling.

Critical strengths

A clear decomposition consistent with deployment: The routing and labeling channels distinguish two roles of provenance that are easily conflated. Their nested structures make source usage a testable choice rather than an unconditional extra input. Keeping the deployed classifier X-only is also appropriate. I view this decomposition and its operational model-selection interpretation as the main conceptual contribution. S2. Matched interventions support source-dependent labeling: Table 6 compares all four forms within the same learning framework: on dopanim, R1L2 achieves 68.81% balanced accuracy versus 67.16% for provenance-blind R1L1. The true-versus-shuffled provenance contrast in Table 3, 96.17% versus 94.04%, further tests whether the source–error correspondence matters. Reporting that CIFAR-10N does not require provenance, and that unsupported routing can hurt, appropriately qualifies the claim.

Critical weaknesses
  1. The operational statistics replace the true class and full input with predicted classes and finite feature strata (Equations 16–18). Coarse conditioning can preserve residual input-related associations, while proxy errors and sparse source–class cells can affect the tests. The authors correctly call these model-relative diagnostics. However, Table 1 contains only three real settings with proxy/oracle decision agreement and no routing-positive case. Replacing the proxy class with the true class still retains representation-based stratification, so that audit is not an independent certification of population conditional independence. Stability across representation quality, stratification, and source density remains unclear, directly limiting the structure-selection contribution.
  2. The best CIFAR-10N model is provenance-blind, so its substantial improvement over DivideMix in Table 4 cannot be attributed to provenance. Positive real-data evidence for provenance is concentrated in the dopanim labeling channel, while routing interactions are small. The paper identifies InstanceGM as its provenance-blind ancestor and provides a variational-branch decomposition, but the implementation-level correspondence and the channels’ incremental value under matched computation should be clearer.
Questions for authors

Q1. On the existing controlled datasets, how do false-positive rates, power, and selection frequencies change across warm-up seeds, numbers of feature strata, and minimum cell sizes? Please report routing-positive cases and distinguish failure to reject from affirmative evidence of independence. Q2. Can you provide seed-level paired R1L2-versus-R1L1 differences on dopanim and cross-seed shuffled-provenance results? Please also state whether provenance-blind PCVL and a matched InstanceGM implementation are identical in objective, architecture, and training configuration.

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

This paper studies noisy-label learning when the source of each label is available during training. It proposes PCVL, which models how examples reach sources and how sources produce labels, together with a permutation diagnostic for deciding which source dependencies to use. The controlled experiments provide useful evidence for source-dependent labeling. However, I am not fully convinced that the diagnostic reliably selects the intended dependencies, and the additional contribution over existing generative noisy-label methods is not sufficiently established.

Critical strengths
  • The problem is practically relevant. Keeping the source of a noisy label is a reasonable way to use information that is often discarded, while still deploying an input-only classifier.
  • The controlled data-model cross and true-versus-shuffled provenance experiment are useful. They help separate the benefit of informative source identity from simply adding another input to the model.
  • The paper is clear about its limitations, including the lack of an identifiability guarantee and the weaker evidence for routing. The matched model variants and loss ablations are also helpful.
Critical weaknesses
  • Experiment evaluation only uses the CIFAR dataset, it would be better if tests can done on more diverse datasets and on real datasets to verify the generalizability of the method.
  • The diagnostic is central to the contribution, but its reliability is not sufficiently established. In Section 3.4, the tests condition on 20 feature clusters and predicted labels rather than the actual input and clean label. Residual variation within a cluster can create an apparent source dependency even when the stated conditional independence holds. The oracle audit still uses the same coarse conditioning, so agreement with it does not resolve this issue. I appreciate the model-relative interpretation, but the controlled experiments should also report the diagnostic's selections, false-positive rates, and detection rates.
  • The provenance-specific benefit is narrower than the main performance table suggests. Both CIFAR-10N settings select the provenance-blind model, so their gains over DivideMix do not demonstrate a benefit from modeling provenance. On real data, the isolated source-dependent labeling gain is 1.65 balanced-accuracy points on dopanim (Table 6). Routing is unsupported in all real-data settings and its controlled effects are small. More evidence is needed for the full two-channel proposal.
  • The distinction from the closest generative noisy-label methods needs to be clearer. The paper identifies InstanceGM as the provenance-free specialization, but does not clearly quantify which implementation or objective changes separate it from PCVL-R1L1. A matched comparison or an explicit equivalence would help explain the large provenance-blind gains and assess what the new framework adds.
Questions for authors
  • can the authors add more datasets and real datasets?
  • Can the authors report diagnostic selections across repeated controlled datasets, including sensitivity to the number of feature clusters and warm-up quality?
  • Is PCVL-R1L1 an exact implementation of InstanceGM? If not, which changes account for its gains, and how does it compare with the original method under the same protocol?
  • Can the source-dependent labeling gain be replicated on another real dataset? Can the shuffled-provenance stress test report uncertainty across seeds rather than only across epochs?
Reviewer_RfwpOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Label Provenance Learning (LPL). The training data is represented as {(xi,si,ai)}\{(x_i, s_i, a_i)\}, where sis_i denotes a potentially noisy observed label, aia_i represents the label provenance (source), and the clean label yy is treated as a latent variable. The authors decompose the role of provenance into two distinct channels: a routing channel RR (governing how samples are assigned to sources) and a labeling channel LL (governing how sources generate labels). Methodologically, the paper proposes Provenance-Conditioned Variational Learning (PCVL), which integrates the Evidence Lower Bound (ELBO) with the sample-selection objective from DivideMix. Additionally, the authors introduce a hierarchical permutation diagnostic to assess which dependencies are supported by the data, without requiring access to clean labels, thereby determining which channels should be activated. The approach is validated through experiments on CIFAR-10N, Dopanim, controlled CIFAR-10 datasets, and "shuffled-provenance" intervention experiments.

Critical strengths
  1. The research problem addressed in this paper is highly relevant to practical applications. Large-scale datasets often contain labels originating from a variety of sources, such as different websites, protocols, or annotators. Standard Learning with Noisy Labels (LNL) approaches typically pool these labels into an undifferentiated set (X,S)(X, S), thereby discarding information about the data-collection process, which could potentially be valuable.
  2. The paper explicitly separates two potential roles of provenance—how samples reach a source versus how a source generates labels—and models them as testable hypotheses, rather than assuming that source information is inherently useful. This design is valuable, with the four resulting configurations serving both as diagnostic indicators and as ablation controls.
  3. Proxy-based diagnosis (using a proxy classifier trained on noisy labels) yields decisions consistent with those from an oracle audit (which uses true, clean labels) across all three real-world datasets (Table 1), providing strong validation for the practical utility of this diagnostic approach.
Critical weaknesses
  1. Compared to DivideMix, PCVL incorporates a decoder and multiple independent encoders; however, its computational overhead and training stability have not been reported.
  2. The author states that InstanceGM is a special case, or an ancestor of (R1, L1). Furthermore, the variant of (R1, L1) selected by PCVL for use on CIFAR-10N is specifically "provenance-blind." However, the experiments lack a comparison with the InstanceGM method.
Questions for authors
  1. DivideMix achieves only 78.15% on the CIFAR-10N top-5 setting, which is significantly lower than the results typically reported for DivideMix on this benchmark in the literature. What is the reason for this?
  2. On the controlled dataset, the full R1L1 method outperforms DivideMix by only 2.08 points, whereas on CIFAR-10N, it shows a lead of 10.44 points (in terms of top-5 accuracy). Please explain this significant disparity in the magnitude of the performance gap.
  3. Why choose λX,λS,λRλ_X, λ_S, λ_R as (1, 1, 1) and βY,βZβ_Y , β_Z = (0, 0)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.