- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper targets the confirmation bias and error accumulation issues caused by candidate label ambiguity in partial label learning, and proposes the DC-RPLL framework. The method estimates sample reliability using neighborhood information, filters samples from easy to hard via dynamic curriculum learning, then leverages Actor-Critic reinforcement learning to independently perform three label disambiguation actions for each sample, and designs a self-supervised reward function to realize policy optimization without ground-truth labels. The proposed method achieves better performance than multiple SOTA algorithms in experiments on benchmark datasets and real partial label datasets, and ablation experiments confirm that each module brings performance gains. However, this work lacks complete theoretical analysis, the model has non-negligible computational complexity, and further verification on larger-scale datasets is not conducted.
Critical strengths
-
Combining RL with sample reliability-aware curriculum learning to solve error propagation in partial label learning. Most existing partial label disambiguation methods update labels following static fixed rules, which easily amplifies errors generated in early training stages. This paper makes adaptive decisions of update, retain and abstain at the individual sample granularity, so as to dynamically avoid noisy label contamination introduced by low-reliability samples.
-
Designing a ground-truth-free self-supervised reward function for reinforcement learning policy training. Since no valid ground-truth labels are available in partial label scenarios, this reward function constructs supervision signals from four dimensions: confidence gain, entropy variation, neighborhood reliability and label drift, enabling the reinforcement learning paradigm to adapt to the weakly supervised setting.
Critical weaknesses
-
The theoretical analysis in this work is relatively weak. It only briefly mentions the error contraction property, and lacks rigorous derivations and proofs for the convergence of the reinforcement learning policy and the effectiveness of the self-supervised reward function. The core innovation of this paper lies in reinforcement learning-driven selective label disambiguation, but it does not fully demonstrate whether the policy can converge stably, or why the reward function can guide the model to obtain correct disambiguation results. The conclusions are only supported by experimental phenomena, which undermines the credibility of the method and stands as a critical shortcoming that affects the overall evaluation.
-
This framework needs to maintain neighborhood relationships and execute AC decisions for each individual sample, which leads to high time complexity. The paper only provides the complexity formula, without offering efficiency comparison results or feasible schemes to alleviate the computational burden.
Questions for authors
-
The paper mentions that the proposed reliability-aware correction has the error contraction property, but the appendix only provides the objective function analysis, without corresponding complete derivations. Could you supplement and specify under what exact assumptions this error contraction can be guaranteed to hold? If these assumptions cannot be satisfied, will DC-RPLL suffer from label disambiguation failure?
-
The self-supervised reward is the core of the entire reinforcement learning policy, and different hyperparameters will alter the weight of each reward term. Beyond the parameter sensitivity experiments presented in the paper, have you verified that when the reward weights are drastically perturbed, whether the policy will degenerate to always choosing abstention or performing blind updates, which further leads to severe performance degradation?
4 comments
Q2:
Thank you for asking whether large perturbations to the self-supervised reward can cause persistent abstention or indiscriminate updates. To extend the sensitivity analysis in Fig. 4, we conducted a one-factor-at-a-time stress test on CIFAR-10 with . We varied the entropy-reduction weight , the reliability reward weight , and the label-drift penalty weight over and times their default values of , , and , respectively. The three sweeps share the same default configuration, giving 13 configurations in total. Each perturbation preserves the reward formulation and label-correction operator.
| Configuration | Best test accuracy | Update coverage |
|---|---|---|
| DC-RPLL (default) | 89.15% | 95.62% |
| 87.91% | 92.50% | |
| 88.87% | 92.76% | |
| 88.41% | 92.27% | |
| 87.47% | 92.80% | |
| 86.91% | 92.41% | |
| 88.40% | 92.89% | |
| 87.10% | 92.76% | |
| 85.99% | 92.54% | |
| 87.54% | 92.75% | |
| 88.21% | 93.10% | |
| 87.44% | 92.60% | |
| 87.32% | 92.30% |
Based on these results, we can explicitly address the reviewer's two concerns:
- No Degeneration into "Always Abstaining": If the policy degenerated into always choosing abstention, the update coverage would plummet to near zero. However, as shown in the table, across all extreme perturbation settings,including completely zeroing out the entropy or reliability rewards,the update coverage remains highly stable, consistently staying above 92% (compared to 95.62% for the default). This demonstrates that the policy does not over-correct into a state of extreme conservatism (abstention).
- No Degeneration into "Blind Updates": If the policy engaged in blind updates, the introduction of excessive incorrect pseudo-labels would cause the accuracy to catastrophically collapse (e.g., dropping to near random guess or baseline levels without RL). While the accuracy does experience a controlled degradation under severe weight mismatch (e.g., dropping to 85.99% when ), it remains highly competitive and stable. The worst-case accuracy in this extreme perturbation matrix is 85.99%, which is still far above the random baseline and comparable to the non-learning contextual bandit baseline (86.11%) reported earlier. This indicates that the policy maintains its discriminative capability and does not blindly destroy the label space.
Over-emphasizing reliability () causes the most noticeable performance drop (85.99%), as it likely forces corrections on samples that are locally consistent but globally misclassified. Completely removing the entropy reward () degrades accuracy to 87.91%, proving its necessity in preventing premature convergence, yet the system remains stable. The distribution-drift penalty () shows the highest tolerance to perturbation, with accuracy remaining robust across all scales. The extreme perturbation study confirms that while optimal reward balancing yields the highest accuracy (89.15%), the policy is highly robust against severe reward weight misspecification. It does not suffer from catastrophic collapse into permanent abstention or blind updating.
Discussion
Thanks for your response.
Thank you for taking the time to review our response. We are happy to clarify any remaining questions. Best regards, The authors
We thank the reviewer for the careful review and valuable feedback. We address the concerns below.
We thank the reviewer for the careful review and valuable feedback. We address the concerns below.
Q1:
Thank you for identifying this gap. We have revised the appendix to provide complete derivations and explicit sufficient assumptions. The algorithm, reward function, correction target, and normalization order remain unchanged.
For the original affine correction, let Ui(e) indicate an update action and define the current and proposal errors as di(e)=∥qi(e)−eyi∥1 and d~i(e)=∥q~i(e)−eyi∥1. The identity ∥q−ey∥1=2(1−q(y)) gives
=d_i^{(e)}+ U_i^{(e)}\beta_i^{(e)} (\tilde d_i^{(e)}-d_i^{(e)}).$$ Thus, a positive-step update reduces error exactly when the proposal assigns greater probability to the true class than the current label. Expected contraction requires: (i) feasible probability distributions with the stated candidate-support conditions; (ii) selected-proposal quality, $$\mathbb E[\tilde d_i^{(e)} \mid\mathcal F_e,U_i^{(e)}=1,\beta_i^{(e)}] \le\kappa d_i^{(e)}+\epsilon,\quad \kappa<1;$$ and (iii) effective update opportunities bounded by $$0<\omega_{\min}\le \mathbb E[U_i^{(e)}\beta_i^{(e)}\mid\mathcal F_e] \le\omega_{\max}.$$ These yield $$\mathbb E[d_i^{(e)}] \le\rho^e d_i^{(0)} +\frac{\omega_{\max}\epsilon}{1-\rho}(1-\rho^e), \qquad \rho=1-(1-\kappa)\omega_{\min}<1.$$ For the original operator that normalizes **after interpolation**, we preserve this order and derive its exact effective step: $$\lambda_i=\frac{\beta_i m_i}{1-\beta_i+\beta_i m_i}, \qquad m_i=\sum_{c\in S_i}\tilde q_i(c).$$ The corresponding proof uses $\lambda_i$ and the analytically conditioned proposal. We additionally relate proposal error to cross-class attention mass and same-class neighbors’ soft-label errors. High reliability alone does not establish these conditions. Zero residual error permits geometric contraction; persistent proposal error generally permits only a residual-error bound. If the assumptions fail, disambiguation does not necessarily fail, but contraction is no longer guaranteed. Incorrect consensus may reinforce errors, and persistent abstention may prevent recovery. Adaptive steps limit correction magnitude without certifying direction. Therefore, we have revised our claims regarding unconditional calibration and joint convergence, and we explicitly distinguish between the contraction of conditional label errors and the convergence of the overall training algorithm.