Structured Prediction Selective Pseudo-Labeling

• The paper introduces a framework that selectively assigns pseudo-labels based on output structure and model uncertainty to improve training efficiency.
• It employs techniques such as clustering-driven labeling, entropy partitioning, and ensemble-based estimation to partition reliable and ambiguous predictions.
• Empirical results demonstrate that this approach enhances performance in tasks like semantic segmentation and NER while reducing annotation costs.

Structured prediction based selective pseudo-labeling is a family of learning paradigms that address the challenge of leveraging abundant unlabeled data in structured output problems by carefully selecting and utilizing pseudo-labels guided by the structure present in either the output space or the model’s uncertainty. This approach has demonstrated superior data efficiency, robustness to noisy labeling, and improved performance across a range of domains, including semantic segmentation, natural language processing, text recognition, and unsupervised domain adaptation.

1. Conceptual Foundations and Motivation

In structured prediction tasks, the output space 𝒴 is exponentially large and structured, involving interdependent sub-parts such as pixels in segmentation maps, tokens in sequence labeling, or relational edges in parsing. Semi-supervised and unsupervised adaptation methods often rely on pseudo-labeling, where model predictions on unlabeled data are treated as surrogate ground truth. However, naive pseudo-labeling—accepting all or any high-confidence predictions—frequently results in label noise propagation, class imbalance, and degraded sample efficiency, especially under domain shift conditions or for ambiguous/hard samples.

The key insight underlying structured prediction based selective pseudo-labeling is the use of both the inherent structure of the problem and statistical uncertainty to guide which pseudo-labels are used for supervision and how ambiguous cases contribute to training. Selection strategies may rely on clustering in feature space, uncertainty quantification (entropy, margin, dispersion), expert strategies (partial annotation, error prediction), or domain-specific curricula. The aim is to maximize the value of each unlabeled sample in training without amplifying noise or bias (Wang et al., 2019, Wang et al., 2023, Zhang et al., 2023, Liu et al., 20 Sep 2025, Patel et al., 2022, Cho et al., 2023).

2. Structured Prediction and Selective Pseudo-Labeling Frameworks

Multiple instantiations of selective pseudo-labeling exist, each exploiting structure and uncertainty for label selection and utilization:

• Clustering-Driven Labeling: In unsupervised domain adaptation, samples in the target domain are clustered in the deep feature space, allowing for class-balanced selection of high-confidence pseudo-labels matched via prototype-based or cluster-to-class alignment. Fused probability scores from nearest-class and structured-prediction clustering are used to assign pseudo-labels, and only the most confident samples per class are used at each iteration. Progressive expansion ensures class balance and reduces error accumulation (Wang et al., 2019).
• Entropy and Dispersion Partitioning: In dense prediction (e.g., semantic segmentation), per-pixel entropy or confidence-dispersion features partition predictions into “reliable” (high confidence) for positive pseudo-labeling and “unreliable” (ambiguous) pixels, which can be repurposed as negative keys or subjected to adaptive weighting (Wang et al., 2023, Liu et al., 20 Sep 2025). Adaptive quantile thresholds or spectral separation in confidence space enable robust delineation of usable and high-risk regions.
• Partial Annotation and Error-Estimating Selection: For structured tasks such as NER or dependency parsing, partial manual annotation of uncertain sub-structures is guided by an explicit error estimator trained on log-margins. The ratio of selected to pseudo-labeled structures is set to the estimated error rate, ensuring annotation effort is directed to informative but hard regions, with the remainder auto-labeled via the current model’s distribution (Zhang et al., 2023).
• Curriculum-Guided Selective Inclusion: In cell detection, pseudo-heatmaps constructed from local maxima serve as candidate pseudo-labels. Bayesian uncertainty estimation filters candidates, and an easy-to-hard curriculum on cell density limits overfitting to difficult patches early in training (Cho et al., 2023).

3. Uncertainty Quantification and the Reliable/Unreliable Partition

Central to selective pseudo-labeling is robust uncertainty estimation. Techniques include:

• Entropy and Confidence Margin: Shannon entropy ($H(p_{ij}) = -\sum_c p_{ij}(c) \log p_{ij}(c)$) or margin-based uncertainty separate reliable and unreliable predictions (Wang et al., 2023, Zhang et al., 2023).
• Confidence Dispersion and Spectral Analysis: Construction of a 2-D vector of maximum class confidence and residual dispersion forms the basis for spectral separation of pseudo-label eligibility. Spectral relaxation (top eigenvectors of the covariance matrix) yields adaptive boundaries between reliable and unreliable predictions per mini-batch (Liu et al., 20 Sep 2025).
• Ensemble-based Bayesian Estimation: MC-Dropout ensembles or Bayesian CNNs yield predictive distributions whose entropy or variation ratio provide selection criteria for sequence-level or structural pseudo-labels (Patel et al., 2022, Cho et al., 2023).
• Error Modeling and Adaptive Ratios: An error estimator trained on log-margins directly predicts the probability of correctness for sub-structures, from which the partial annotation ratio is adaptively set to match the model’s dynamic error (Zhang et al., 2023).

4. Training Objectives and Pipeline Integration

Selective pseudo-labeling frameworks typically optimize composite objective functions integrating both explicit and implicit supervision:

• Supervised Losses: Fully (or partially) labeled examples yield standard supervised losses; e.g., cross-entropy for classification/segmentation, or log-marginal likelihood for structured outputs.
• Unsupervised/Pseudo-Label Losses: Reliable pseudo-labels are incorporated into the loss (e.g., symmetric cross-entropy $SCE(p,y)=\xi_1 CE(p,y)+\xi_2 CE(y,p)$, where $\xi_2>0$ mitigates label noise effects (Wang et al., 2023)); unreliable or ambiguous predictions are often excluded from direct supervision but used as negative or contrasting samples.
• Contrastive and Negative Key Losses: In semantic segmentation, contrastive losses with category-wise negative queues use embeddings of unreliable pixels to encourage class separation, substantially increasing label efficiency and mitigating confirmation bias (Wang et al., 2023).
• Distillation/Marginal Losses: Distillation of predicted distributions (from older models or ensembles) regularizes the training signal for unlabeled or partially labeled data (Zhang et al., 2023, Patel et al., 2022).
• Random Masking and Smooth Weighting: Trusted mask perturbation or patch-wise smoothing further prevents the model from relying solely on the easy set, helping it learn from ambiguous and contextually valuable regions (Liu et al., 20 Sep 2025).

A typical pipeline involves batch sampling; feature/embedding generation; uncertainty-based partitioning and selection; loss computation integrating labeled, selected, and pseudo-labeled points; model update; and periodic update of error estimators, prototypes, or curriculum parameters. Category-wise or per-class queue structures are maintained for contrastive negative mining (Wang et al., 2023).

5. Empirical Results and Ablation Insights

Selective pseudo-labeling consistently outperforms baseline and contemporary approaches across tasks and domains:

• Semantic Segmentation: On VOC ‘12 and Cityscapes, methods such as U²PL+ achieve 1–4 mIoU improvements over strong alternatives. Dynamic entropy partitioning and category-wise negative queues are crucial for gains; removing unreliable pixels or using only low-entropy “negatives” results in 2–3 mIoU loss. Adjusting $\alpha_t$ and symmetric cross-entropy each provide measurable improvements (Wang et al., 2023).
• Domain Adaptation: For source→target transfers (e.g., GTA5→Cityscapes; various cross-culture cell detection), structured and selective pseudo-labeling methods show clear superiority, yielding 1–1.5+ mIoU or significant F₁ improvements over prior methods and ablations (Wang et al., 2019, Cho et al., 2023).
• Structured Sequence Tasks: In NER, dependency, and event extraction, adaptive partial annotation with selective pseudo-labeling matches full-annotation performance at less than half the reading/labeling cost (Zhang et al., 2023).
• Text Recognition: Seq-UPS delivers substantial word accuracy and CER gains by combining dropout-based uncertainty, beam search candidate generation, and masked selection (Patel et al., 2022).

Ablation studies highlight the deleterious effect of fixed thresholds, motivate the inclusion of ambiguous samples as negatives or context anchors, and reinforce the importance of error-adaptive, class-balanced selection.

6. Comparison of Approaches

A summary of core methodologies and their distinguishing features:

Method (Paper)	Selection Mechanism	Key Use of Structure
U²PL+ (Wang et al., 2023)	Entropy quantile, negative queues	Per-pixel class-wise queues, contrastive embedding
CSL (Liu et al., 20 Sep 2025)	Spectral clustering in (conf, disp) space	Adaptive, smooth, and random mask weighting
SPL (Wang et al., 2019)	Clustering and class prototype matching	K-means, class-balanced selection
Data-efficient Active Learning (Zhang et al., 2023)	Error model + adaptive annotation ratio	Partial annotation and sub-structure selection
Seq-UPS (Patel et al., 2022)	MC-dropout uncertainty on sequences	Sequence/char-level and ensemble selection
Cell Detection (Cho et al., 2023)	Bayesian classifier uncertainty/curriculum	Heatmap regeneration and curriculum on cell density

Each advances the field by fusing uncertainty/risk estimation with explicit exploitation of the structured nature of the output space.

7. Limitations, Open Problems, and Future Directions

Despite their effectiveness, several methodological and computational constraints remain:

• Scalability: Some approaches rely on O(n²) storage for affinity/similarity or spectral matrices, constraining scalability. Potential improvements include kernel approximations, subsampling, or landmark-based methods (Wang et al., 2019).
• End-to-End Deep Learning Integration: Many structured pseudo-labeling algorithms operate on extracted or pre-trained features. Integrating these with fully end-to-end trainable architectures opens further performance and representation learning benefits (Wang et al., 2019).
• Selection Criterion Tuning: Model and task-specific tuning of thresholds, mask smoothing parameters, and negative sample definitions remains necessary, and automated or learnable selection criteria are an ongoing research area.
• Cluster Number and Structure Discovery: Fixed or poorly matched cluster numbers can limit label purity; improved clustering or hierarchical models may enhance pseudo-label reliability (Wang et al., 2019).

The body of evidence demonstrates that leveraging structure in both the data and output space, combined with rigorous uncertainty-aware selection, fundamentally improves the reliability and efficiency of pseudo-labeling in structured prediction tasks (Wang et al., 2023, Liu et al., 20 Sep 2025, Zhang et al., 2023, Patel et al., 2022, Cho et al., 2023, Wang et al., 2019).