Controllable Pseudo-label Generation (CPG)

• Controllable Pseudo-label Generation (CPG) is a framework that refines pseudo-labels through iterative selection and dynamic filtering to mitigate overconfidence, class imbalance, and label noise.
• It employs multi-view agreement, feature-space optimization, and auxiliary unsupervised branches to enhance reliability and enforce desired class distributions.
• Empirical results demonstrate that CPG improves accuracy and calibration in semi-supervised learning, domain adaptation, and long-tailed settings across various benchmarks.

Controllable Pseudo-label Generation (CPG) is a paradigm that systematically refines and regulates the creation of pseudo-labels for unlabeled data, enabling robust semi-supervised learning, domain adaptation, and generative modeling under heterogeneous or uncertain data distributions. CPG frameworks incorporate iterative selection, dynamic filtering, optimization-based confidence thresholding, and network-driven supervision mechanisms to mitigate issues such as overconfidence, class imbalance, confirmation bias, and structured label noise. The overarching goal is to expand the labeled training set with highly reliable pseudo-labeled examples, ensuring statistical or semantic control over the evolving “effective” training distribution, while maximizing data utilization across both labeled and unlabeled sets.

1. Theoretical Foundations and Motivation

CPG arises in contexts where labeled data is limited or long-tailed, and the unlabeled dataset distribution is unknown or arbitrary. Traditional self-training and pseudo-labeling suffer from confirmation bias, unreliable confidence scores, and propensity for majority-class reinforcement. CPG methods depart from static confidence selection schemes by introducing principled mechanisms to control which unlabeled samples are pseudo-labeled and added to the training set, typically via dynamic filtering, agreement-disagreement protocols, optimization in feature-confidence space, or reference-based assignment.

In frameworks such as Confidence Separable Learning (CSL) (Liu et al., 20 Sep 2025), pseudo-label selection is formulated as a convex optimization problem over a multi-dimensional confidence feature space, where sample-specific decision boundaries are established using spectral relaxation. In the context of long-tailed semi-supervised learning (Hou et al., 5 Oct 2025), CPG leverages an iterative controllable self-reinforcing optimization cycle—they dynamically filter reliable pseudo-labels, update the labeled set to enforce a known class distribution, and build a Bayes-optimal classifier using logit adjustment tailored to the current distribution. Theoretically, these cycles are shown to reduce generalization error provided the cumulative risk reduction outpaces pseudo-label noise effects.

2. Dynamic Filtering and Reliability Estimation

Central to CPG is the mechanism by which unlabeled samples are selected and pseudo-labeled. Rather than employing fixed thresholds on softmax confidence, CPG mechanisms introduce additional reliability criteria:

• Multi-view agreement: In (Hou et al., 5 Oct 2025), an unlabeled sample is pseudo-labeled only if both weak and strong augmented views yield the same prediction and their confidence scores exceed a threshold, further corroborated by multi-iteration voting strategies.
• Feature-space optimization: CSL (Liu et al., 20 Sep 2025) defines for each pixel a feature vector $h_n = [p_n(k^*), v_n]^T$, where $p_n(k^*)$ is the maximum predicted class probability and $v_n$ quantifies residual dispersion. Pseudo-label selection is cast as maximizing $\mathrm{Tr}(S^T \Phi^T\Phi S)$, where $S$ is a selection matrix, yielding sample-specific optimal thresholds that account for model overconfidence and dispersion in probabilistic outputs.
• Agreement/disagreement duality: Conservative-Progressive Collaborative Learning (CPCL) (Fan et al., 2022) simultaneously trains two networks and utilizes their agreement for intersection supervision (high-reliability labels) and their disagreement for union supervision (broader exploration via class-wise resolution), with confidence-based dynamic loss weighting suppressing noisy labels.

These mechanisms enable fine-grained control over pseudo-label incorporation, balancing the need for high quality labels with the imperative to leverage large unlabeled datasets.

3. Class Distribution Control and Bayesian Adjustment

A distinguishing attribute of CPG is its ability to enforce or exploit a known or desired class distribution within the effective training set.

• Label distribution tracking: As reliable pseudo-labels are accumulated (filtered via dynamic mechanisms), their class frequencies $\phi_c$ are tracked and normalized to form an empirical priors vector $\pi_c$. In CPG (Hou et al., 5 Oct 2025), the classifier is then retrained using logit adjustment with these priors: for class $y$, logits are modified by adding $\ln P_y$, yielding the logit-adjusted softmax loss

$$\ell_{la}(y, f(g(x))) = -\log\left( \frac{\exp(f_y(g(x)) + \ln P_y)}{\sum_{y'} \exp(f_{y'}(g(x)) + \ln P_{y'})} \right)$$

where $f_y(g(x))$ is the output logit, $g(\cdot)$ is the encoder, and $P_y$ is the class prior. This ensures Bayes-optimality for the controlled empirical distribution.

• Balanced candidate label generation: Candidate Pseudolabel Learning (CPL) (Zhang et al., 15 Jun 2024) proceeds by intra-instance selection (top labels whose cumulative confidence exceeds an adaptively chosen threshold $\tau$) and inter-instance selection (retaining classes whose confidence exceeds the class-wise quantile $\beta$), producing final candidate sets via their intersection. This adaptive mechanism maximizes inclusion of true labels and enforces class balance, leading to improved model calibration.

By modulating the evolving labeled set to match desired priors, CPG decouples training dynamics from the arbitrary or imbalanced distribution of the unlabeled pool, reducing the risk of majority-class bias amplification and supporting learning in realistic, distribution-shifted scenarios.

4. Auxiliary Modules for Minority Classes and Data Utilization

CPG frameworks integrate modules to mitigate the challenges of class imbalance and limited diversity in minority classes:

• Class-aware adaptive augmentation: In (Hou et al., 5 Oct 2025), representations for minority class samples are perturbed via additive noise scaled inversely with class compactness $\alpha(c)$, where $r(c) = 1/\alpha(c)$ and compactness is the average cosine similarity to the class centroid. This adaptive strategy sharpens decision boundaries and increases representation diversity for underrepresented classes.
• Auxiliary branches: To leverage all available samples—including those failing strict reliability criteria—an auxiliary branch is trained via unsupervised consistency regularization on weak-strong augmented pairs, ensuring that representational learning benefits from the full data pool. The overall loss combines logit-adjusted supervised loss and auxiliary unsupervised loss, modulated by an indicator parameter for branch-specific weighting.

These additions further stabilize training and maximize the utility of both labeled and unlabeled data, especially early in the optimization cycle when strict filtering yields few “trustworthy” pseudo-labels.

5. Comparative Performance and Empirical Results

CPG frameworks, as demonstrated on benchmarks such as CIFAR-10-LT, CIFAR-100-LT, Food-101-LT, SVHN-LT, PASCAL VOC, Cityscapes, and MS-COCO, consistently outperform prior state-of-the-art across settings characterized by class imbalance, non-uniform or unknown unlabeled distributions, and low labeled data regimes.

• On CIFAR-10-LT with arbitrary unlabeled distributions, CPG (Hou et al., 5 Oct 2025) improves accuracy by up to 15.97 percentage points over previous methods and achieves an average improvement of 11.14pp across benchmarks.
• CSL (Liu et al., 20 Sep 2025) reports +31.4% mIoU over the supervised baseline using only 1/16th labeled images and competitive performance on Cityscapes and MS-COCO.
• CPL (Zhang et al., 15 Jun 2024) yields higher true label inclusion rates (e.g., ~85% for candidate sets vs. ~70% for hard labels) and improved class balance, with accuracy gains of 10–18pp over GRIP and FPL.

These empirical results validate the effectiveness of CPG’s controlled selection, distribution-enforcing, and auxiliary mechanisms for pseudo-label generation in both vision and multimodal settings.

6. Connections to Related Research and Future Directions

CPG connects to a spectrum of recent advances: iterative generative refinement with cGANs (Morerio et al., 2020), unsupervised conditional clustering (Noroozi, 2020), reference-guided pseudo-labeling (Seibold et al., 2021), collaborative agreement-disagreement learning (Fan et al., 2022), and noisy dual self-training for text generation (Feng et al., 2022). Across modalities, CPG strategies systematically mitigate confirmation bias, optimize label quality (via feedback loops, clustering, or controlled confidence), and adapt to realistic data heterogeneity.

Ongoing questions concern scalability to high-resolution and large-scale data, parameter sensitivity for candidate selection, efficient memory and computational strategies (especially for reference-guided methods and spectral optimization), and integration with direct confidence calibration. Extensions to domain adaptation, general semi-supervised classification, and multimodal generative modeling (e.g., video or cross-modal synthesis) are actively being explored.

A plausible implication is that future CPG designs will incorporate adaptive selection criteria, dynamic thresholding, and class-aware calibrations at scale, potentially supplemented by self-supervised or contrastive representations for improved robustness in additional domains.

7. Practical Availability and Implementation Resources

The reproducibility and deployment of CPG frameworks are facilitated by open-source implementations:

• CSL (Liu et al., 20 Sep 2025): https://github.com/PanLiuCSU/CSL
• CPG for long-tailed SSL (Hou et al., 5 Oct 2025): https://github.com/yaxinhou/CPG
• CPL (Zhang et al., 15 Jun 2024): https://github.com/vanillaer/CPL-ICML2024

These repositories provide scripts for constructing confidence matrices, implementing dynamic selection and filtering, configuring prior distributions and logit adjustment losses, and integrating auxiliary unsupervised branches. Recommended hyperparameter settings and experiment instructions support rapid adoption and benchmarking within existing semi-supervised and domain-adapted pipelines.

---

CPG formalizes the principle of controlling pseudo-label generation via dynamic filtering, empirical distribution enforcement, and tailored auxiliary strategies. Its data-driven, optimization-based frameworks substantially improve learning performance in semi-supervised, domain-adapted, and long-tailed settings characterized by heterogeneity, class imbalance, and unreliable confidence signals. The approach is poised to underpin future developments in robust machine learning under practical, real-world data scenarios.