Uncertainty-aware Pseudo-label Generation

• The paper shows that incorporating uncertainty into pseudo-label generation effectively mitigates overconfident errors and noisy predictions.
• UPG is a design principle that selects, refines, or weights pseudo-labels based on explicit uncertainty signals, enhancing model reliability across domains.
• The approach uses techniques like MC-Dropout, entropy estimation, and ensemble consistency to balance pseudo-label quality with computational costs.

Uncertainty-aware Pseudo-label Generation (UPG) denotes a family of self-training and adaptation procedures in which pseudo-labels are not treated as uniformly trustworthy, but are generated, selected, refined, or weighted according to an explicit uncertainty signal. Across semi-supervised learning, unsupervised domain adaptation, source-free adaptation, dense prediction, sequence modeling, graph learning, and stereo matching, the recurring motivation is the same: pseudo-labeling fails when models produce incorrect predictions with high confidence, when uncertainty concentrates at boundaries or ambiguous regions, or when structured outputs accumulate autoregressive or spatial errors (Hegde et al., 2021, Cai et al., 2021, Patel et al., 2022, Litrico et al., 2023).

1. Historical and conceptual scope

UPG emerged as a response to a widely reported pathology of pseudo-label-based learning: the model trains on its own predictions, but those predictions are often noisy, miscalibrated, or domain-shifted. In semi-supervised classification, this was formalized as overconfident wrong predictions from poorly calibrated models, which make confidence-only thresholding unreliable (Rizve et al., 2021). In source-free unsupervised domain adaptation for 3D object detection, the same issue appears as incorrect source-model pseudo-labels with high confidence, rendering thresholding ineffective (Hegde et al., 2021). In semi-supervised text recognition, the problem is compounded by combinatorial sequence spaces and error accumulation across autoregressive decoding steps (Patel et al., 2022). In medical segmentation, pseudo-label errors are reported to concentrate in high-uncertainty regions, especially around object boundaries (Zhou et al., 11 Feb 2025).

The term UPG therefore covers more than one operation. In some papers it refers to uncertainty-aware pseudo-label selection, as in UPS for semi-supervised learning (Rizve et al., 2021) and Bayesian Faster R-CNN-based cross-domain object detection (Cai et al., 2021). In others it refers to pseudo-label refinement by neighborhood aggregation or EM-style posterior updates, as in source-free classification with entropy-weighted neighborhood voting (Litrico et al., 2023) and graph self-training with soft posteriors $Q(\mathbf{z}_v=k)=p_\theta(y=k\mid v)$ (Wang et al., 26 Mar 2025). In dense prediction, it often means uncertainty-aware utilization of pseudo-labels, with hard supervision in reliable regions and soft or attenuated supervision elsewhere (Gröger et al., 2021, Li et al., 8 May 2025, Yan et al., 2022).

A persistent conceptual distinction in the literature is that uncertainty may be used either before pseudo-label formation, during pseudo-label refinement, or after label assignment as a reliability weight. This suggests that UPG is best understood as a design principle rather than a single algorithmic template.

2. Uncertainty estimators and pseudo-label representations

The literature uses several distinct uncertainty estimators. A common family is Bayesian or ensemble-style predictive uncertainty. UPS uses MC-Dropout with 10 stochastic forward passes and standard deviation of output probabilities as the uncertainty score (Rizve et al., 2021). STRUDEL estimates pixel-wise uncertainty as MC-dropout variance over $C=10$ stochastic passes with dropout rate $0.2$ (Gröger et al., 2021). Seq-UPS uses MC-Dropout with dropout probability $p=0.1$ and ensemble size $K=5$ to estimate uncertainty over beam-search hypotheses in autoregressive text recognition (Patel et al., 2022). In unsupervised cross-domain object detection, a Bayesian Faster R-CNN is constructed via dropout as approximate variational inference, and pseudo-label generation is built on that uncertainty model (Cai et al., 2021).

A second family uses entropy or posterior sharpness. In source-free classification, neighborhood-refined pseudo-labels are derived from

$$\hat p_t^{(c)}=\frac{1}{K}\sum_{i\in\mathcal I} p_i'^{(c)}, \qquad \hat y_t=\arg\max_c \hat p_t^{(c)},$$

and uncertainty is measured by normalized entropy

$$\hat{\mathcal H}(\hat p_t)=\frac{-\sum_{c=1}^C \hat p_t^c \log_2 \hat p_t^c}{\log_2 C}, \qquad w_{x_t}= \exp\!\big(-\hat{\mathcal H}(\hat p_t)\big)$$

(Litrico et al., 2023). In graph self-training, the uncertainty signal is simply the softness of the posterior $Q(\mathbf{z}_v)$, with hard pseudo-labeling only when $\max_k Q(\mathbf{z}_v=k)>\gamma$ (Wang et al., 26 Mar 2025). In building damage mapping, uncertainty is operationalized by normalized prediction entropy, with pixels satisfying $E_n<0.3$ regarded as reliable (Li et al., 8 May 2025). nnFilterMatch likewise uses Shannon entropy to rank unlabeled samples for filtering inside nnU-Net training (Yang, 24 Sep 2025).

A third family uses prediction consistency. In UDASOD, uncertainty is the variance of saliency predictions under reversible augmentations such as horizontal flipping, rescaling to $C=10$0, and FDA style swapping: $C=10$1 (Yan et al., 2022). In domain-adaptive person re-identification, pseudo-label credibility is estimated from teacher-student consistency in a reference-center space via

$C=10$2

although that method is explicitly a post-assignment reweighting framework rather than pseudo-label generation in the strict sense (Zheng et al., 2020).

A fourth family uses task-specific structured uncertainty. Seq-UPS estimates uncertainty over beam-search sequence hypotheses rather than single-token confidences (Patel et al., 2022). Robust stereo matching uses pixel-level disparity-distribution variance

$C=10$3

and a learned area-level uncertainty network to generate sparse reliable disparity pseudo-labels (Shen et al., 2023). Self-loop uncertainty for medical segmentation uses recurrent self-supervised encoder updates over Jigsaw transformations, then aggregates segmentation predictions with weights

$C=10$4

to approximate ensemble-like uncertainty at lower cost (Li et al., 2020).

Pseudo-label representations also vary. Some methods use hard labels, such as argmax pseudo-label maps in nnFilterMatch (Yang, 24 Sep 2025) or accepted graph node labels $C=10$5 after thresholding (Wang et al., 26 Mar 2025). Some use soft labels, such as self-loop uncertainty maps $C=10$6 (Li et al., 2020) and saliency pseudo-labels $C=10$7 (Yan et al., 2022). Some explicitly combine hard and soft targets: building damage mapping uses hard pseudo-labels for low-uncertainty pixels and KL-based soft supervision for uncertain pixels (Li et al., 8 May 2025). UPS additionally generalizes pseudo-labeling to include negative pseudo-labels for class absence (Rizve et al., 2021).

3. Core algorithmic patterns

A central pattern is hard uncertainty-aware selection. UPS accepts a classwise pseudo-label only if it is both confident and low-uncertainty: $C=10$8 (Rizve et al., 2021). UDTS follows the same logic in long-tailed semi-supervised learning, using adaptive uncertainty and confidence thresholds that change over training and across classes (Yang et al., 2024). In graph learning, nodes with $C=10$9 remain unlabeled, so ambiguous nodes are delayed until later EM iterations (Wang et al., 26 Mar 2025).

An equally important pattern is soft reweighting rather than rejection. In the uncertainty-aware mean teacher for source-free 3D object detection, uncertain pseudo-label losses are downweighted rather than removed, yielding automatic soft-sampling of pseudo-labeled data (Hegde et al., 2021). STRUDEL uses uncertainty-aware BCE

$0.2$0

so highly uncertain pseudo-labeled pixels contribute less to training (Gröger et al., 2021). UDASOD uses pixel-wise weights

$0.2$1

while separately selecting low-uncertainty images at the sample level (Yan et al., 2022). Source-free classification applies a smooth exponential reliability weight $0.2$2, and reports that this soft weighting outperforms hard entropy-threshold filtering (Litrico et al., 2023).

A third pattern is uncertainty-conditioned mixed supervision. In building-guided damage mapping, low-entropy pixels are trained with cross-entropy against hard pseudo-labels, whereas high-entropy pixels are regularized with KL divergence against soft pseudo-labels, using a fixed $0.2$3 weight ratio between the low- and high-uncertainty regions (Li et al., 8 May 2025). Bidirectional uncertainty-aware region learning adopts a complementary asymmetry: on labeled data it emphasizes high-uncertainty regions, and on unlabeled data it emphasizes low-uncertainty regions (Zhou et al., 11 Feb 2025). This directly rejects the idea that uncertainty should always be discarded.

A fourth pattern is iterative refinement of the pseudo-label source itself. Seq-UPS does not pseudo-label from greedy decoding but from beam search over top-$0.2$4 sequence hypotheses (Patel et al., 2022). Source-free classification replaces single-sample predictions with neighborhood-averaged class distributions before computing entropy (Litrico et al., 2023). UES generates pseudo-labels from an uncertainty-weighted multi-head ensemble and maps sample uncertainty to long-tailed utility weights

$0.2$5

thereby avoiding binary accept/reject decisions at the sample level (Wu et al., 13 Mar 2025). This suggests that UPG increasingly treats pseudo-label quality as a graded utility variable rather than a thresholded event.

4. Domain-specific instantiations

In object detection and domain adaptation, UPG is closely tied to domain shift and source bias. Bayesian Faster R-CNN-based unsupervised cross-domain object detection combines feature alignment with uncertainty-aware pseudo-label generation so that output-space supervision is informed by uncertainty estimates rather than raw detector confidence (Cai et al., 2021). In source-free 3D object detection, uncertainty-aware mean teacher learning combines iterative pseudo-label generation with Monte Carlo dropout and soft loss attenuation on KITTI lidar targets (Hegde et al., 2021). In source-free classification, pseudo-labels are progressively refined by target-neighborhood aggregation, then reweighted by entropy-derived reliability, and trained with negative learning and contrastive regularization (Litrico et al., 2023).

In medical and dense image segmentation, UPG is predominantly spatial. STRUDEL couples MC-dropout variance with uncertainty-weighted pseudo-label loss for white matter hyperintensity segmentation, and further bootstraps initial pseudo-labels with LPA via pixel-wise OR (Gröger et al., 2021). Self-loop uncertainty generates weighted soft pseudo-label maps by recurrently updating the encoder on a Jigsaw pretext task, then supervising only pixels with $0.2$6 (Li et al., 2020). Bidirectional uncertainty-aware region learning argues that pseudo-label errors are concentrated in high-uncertainty regions, especially boundaries, and therefore uses opposite uncertainty emphases for labeled and unlabeled branches (Zhou et al., 11 Feb 2025). nnFilterMatch departs from confidence-threshold SSL by retaining dense argmax pseudo-label maps, ranking unlabeled samples by entropy, and removing 15% low-uncertainty samples every 100 epochs to focus training on uncertain, informative cases (Yang, 24 Sep 2025).

In remote sensing and cross-modal damage mapping, pseudo-label reliability is improved upstream through model fusion and test-time augmentation, then used downstream through entropy-based supervision partitioning. Building extraction pseudo-probabilities are constructed as

$0.2$7

with $0.2$8 averaged across multiple PVT-v2 backbones and flip views, after which low-uncertainty pixels satisfy $0.2$9 and receive hard supervision (Li et al., 8 May 2025). The damage stage mirrors this design and additionally uses a building prior to modulate features for damage pseudo-label refinement (Li et al., 8 May 2025).

In sequence and structured prediction, uncertainty is defined over structured outputs rather than independent classes. Seq-UPS argues that raw confidence is weakly correlated with correctness for autoregressive text recognition, uses beam search to obtain pseudo-label candidates, and evaluates uncertainty over both character-level and word-level predictive distributions under MC-Dropout (Patel et al., 2022). In stereo matching, the preferred pseudo-labels are sparse disparity values retained only where pixel-level or area-level uncertainty is below a fixed threshold, because sparse reliable supervision is empirically more effective than dense noisy disparity maps (Shen et al., 2023).

In graphs and classification, the main emphasis is confirmation-bias control. UGST uses EM-style soft posterior estimation followed by max-probability filtering on Cora, Citeseer, and Pubmed (Wang et al., 26 Mar 2025). UPS introduces positive and negative pseudo-labels with MC-Dropout uncertainty gating (Rizve et al., 2021). UDTS adapts uncertainty thresholds dynamically across classes and training stages for long-tailed SSL (Yang et al., 2024). UES replaces fixed confidence thresholds with uncertainty-aware long-tailed sample weights and uncertainty-weighted ensemble heads, integrating into DualPose and FixMatch (Wu et al., 13 Mar 2025).

5. Empirical findings and recurring trade-offs

The empirical record consistently supports the proposition that uncertainty improves pseudo-label quality rather than merely changing optimization dynamics. In UPS on CIFAR-10 with 1000 labels, conventional pseudo-labeling had initial pseudo-label accuracy $p=0.1$0, confidence-based selection reached $p=0.1$1, and UPS initially selected 20217 positive labels with 94.87% accuracy; after 10 pseudo-labeling iterations, it selected about 93.12% of positive pseudo-labels with 90.29% accuracy (Rizve et al., 2021). The same study reports that adding uncertainty on top of confidence-based selection yields a further $p=0.1$2 error reduction on CIFAR-10 with 1000 labels (Rizve et al., 2021).

In source-free classification, the VisDA-C ablation moves from 85.8 to 90.0 when uncertainty reweighting is added on top of refinement, contrastive regularization, negative learning, and temporal negative-pair exclusion (Litrico et al., 2023). That paper also reports that a hard entropy-margin variant reaches 85.9%, while smooth exponential weighting reaches 90.0%, indicating that soft reliability weighting can outperform hard filtering (Litrico et al., 2023).

In graph learning, UGST improves mean accuracy and stability: Cora $p=0.1$3 versus $p=0.1$4 for vanilla self-training, Citeseer $p=0.1$5 versus $p=0.1$6, and Pubmed $p=0.1$7 versus $p=0.1$8, with gains reported up to 2.5% and lower variance across runs (Wang et al., 26 Mar 2025). This is one of the clearest demonstrations that EM-refined soft posteriors plus confidence filtering can stabilize iterative self-training.

In stereo matching, uncertainty-controlled sparsification produces a particularly explicit precision-coverage trade-off. On KITTI2015, UCFNet pretrained only gives 5.2% D1_all, adaptation with pixel-level pseudo-labels gives 3.89%, and adaptation with area-level pseudo-labels gives 3.64% (Shen et al., 2023). Threshold sweeps show that the most accurate pseudo-labels are not necessarily optimal for adaptation: for area-level uncertainty, $p=0.1$9 yields pseudo-label D1_all $K=5$0 at density $K=5$1, but the best adaptation result is 3.64% at $K=5$2, where pseudo-label D1_all is $K=5$3 and density is $K=5$4 (Shen et al., 2023). This directly supports the claim that UPG must balance reliability and coverage.

In building damage mapping, the building-guided variant improves development-phase mIoU from 74.105 to 74.440, with class-wise gains for Intact, Damaged, and Destroyed, and the final competition result is 54.283 mIoU, ranking first (Li et al., 8 May 2025). In nnFilterMatch, ACDC with 5% labeled data improves from 82.96 Dice under supervised learning to 89.56 under SSL, and to 89.56–90.06 with the SSL+AL variant depending on split, indicating that entropy-based filtering provides an additional gain over plain pseudo-labeling (Yang, 24 Sep 2025).

These findings collectively indicate that uncertainty can increase pseudo-label precision, reduce noise propagation, and improve stability. A plausible implication is that the benefit is strongest when uncertainty is aligned with the exact mechanism that generated the pseudo-label—beam hypotheses in text, neighborhood consensus in source-free classification, or cost-volume sharpness in stereo—rather than being added as a generic auxiliary score.

6. Limitations, misconceptions, and open questions

A recurrent limitation is that many methods use practical uncertainty proxies rather than calibrated uncertainty models. UGST explicitly does not introduce Bayesian posteriors, ensembles, MC-Dropout variance, or aleatoric/epistemic decomposition; it interprets posterior sharpness as uncertainty (Wang et al., 26 Mar 2025). The source-free classification method uses neighborhood entropy as a reliability proxy rather than a formal Bayesian estimate (Litrico et al., 2023). Building damage mapping relies on normalized prediction entropy with a fixed threshold $K=5$5 and no explicit calibration step (Li et al., 8 May 2025). nnFilterMatch likewise depends on single-model predictive entropy without teacher models, ensembles, or calibration (Yang, 24 Sep 2025). UES uses disagreement across heads, and its own discussion notes that the uncertainty-to-weight mapping is static and lacks adaptability to different SSL stages (Wu et al., 13 Mar 2025).

Another limitation is threshold and schedule sensitivity. UPS requires $K=5$6, though it reports that $K=5$7 yields similar CIFAR-10 performance (Rizve et al., 2021). UGST depends on $K=5$8 tuned over $K=5$9 (Wang et al., 26 Mar 2025). Stereo adaptation uses empirically chosen thresholds such as $$\hat p_t^{(c)}=\frac{1}{K}\sum_{i\in\mathcal I} p_i'^{(c)}, \qquad \hat y_t=\arg\max_c \hat p_t^{(c)},$$0 for $$\hat p_t^{(c)}=\frac{1}{K}\sum_{i\in\mathcal I} p_i'^{(c)}, \qquad \hat y_t=\arg\max_c \hat p_t^{(c)},$$1 and $$\hat p_t^{(c)}=\frac{1}{K}\sum_{i\in\mathcal I} p_i'^{(c)}, \qquad \hat y_t=\arg\max_c \hat p_t^{(c)},$$2 for $$\hat p_t^{(c)}=\frac{1}{K}\sum_{i\in\mathcal I} p_i'^{(c)}, \qquad \hat y_t=\arg\max_c \hat p_t^{(c)},$$3 (Shen et al., 2023). The remote-sensing framework fixes $$\hat p_t^{(c)}=\frac{1}{K}\sum_{i\in\mathcal I} p_i'^{(c)}, \qquad \hat y_t=\arg\max_c \hat p_t^{(c)},$$4 as the low-uncertainty rule (Li et al., 8 May 2025). These designs work empirically, but the general threshold selection problem remains open.

A common misconception is that uncertainty-aware pseudo-labeling always means discarding uncertain predictions. Several papers argue against this simplification. STRUDEL downweights rather than rejects uncertain pixels (Gröger et al., 2021). BURL explicitly criticizes direct discarding of high-uncertainty regions because this may neglect valuable clinically meaningful pixels (Zhou et al., 11 Feb 2025). The source-free classification study shows soft exponential weighting outperforming hard entropy filtering (Litrico et al., 2023). nnFilterMatch even reverses the usual logic by removing low-uncertainty samples so training focuses on uncertain, informative cases (Yang, 24 Sep 2025). UES goes further by assigning every sample a nonzero long-tailed weight rather than using binary acceptance (Wu et al., 13 Mar 2025). The literature therefore does not support a single doctrine of “use only low uncertainty”; instead, it supports task-dependent uncertainty modulation.

A further limitation is computational overhead. UPS adds 10 stochastic forward passes and retrains a fresh model each round (Rizve et al., 2021). STRUDEL performs MC-dropout inference and iterative retraining from scratch after fine-tuning (Gröger et al., 2021). Seq-UPS combines beam search with dropout-based ensemble estimation (Patel et al., 2022). Stereo adaptation adds an area-level uncertainty network and iterative pseudo-label regeneration (Shen et al., 2023). These costs are often accepted because pseudo-label noise is otherwise damaging, but they remain operational constraints.

Finally, many strong results depend on auxiliary priors or domain-specific structure. STRUDEL benefits substantially from OR-fusing pseudo-labels with LPA, and performance drops from 0.78 ± 0.10 DSC to 0.67 ± 0.20 without LPA for OctSE-Net on ADNI (Gröger et al., 2021). Building-guided damage mapping depends on the quality of stage-1 building priors (Li et al., 8 May 2025). Stereo pseudo-labeling depends on the source-pretrained matcher being strong enough to produce conservative reliable pixels (Shen et al., 2023). This suggests that UPG is often most effective when uncertainty estimation is coupled with task-specific structural priors, but it also means that poor priors can propagate bias.

Taken together, the literature presents UPG as a shift from unconditional self-training toward reliability-aware self-supervision. The most stable design lesson is not that one uncertainty estimator dominates all others, but that pseudo-label quality improves when uncertainty is attached to the generative mechanism of the pseudo-label itself and when the training system can use that uncertainty flexibly—as a selector, a refiner, a weight, or a curriculum signal—rather than as a single hard threshold.