---
title: Uncertainty-aware Pseudo-label Generation
url: https://www.emergentmind.com/topics/uncertainty-aware-pseudo-label-generation-upg
type: topic
---

# Uncertainty-aware Pseudo-label Generation

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 [2109.14651] [2108.12612] [2209.00641] [2303.03770].

## 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 [2101.06329]. 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 [2109.14651]. In semi-supervised text recognition, the problem is compounded by combinatorial sequence spaces and error accumulation across autoregressive decoding steps [2209.00641]. In medical segmentation, pseudo-label errors are reported to concentrate in high-uncertainty regions, especially around object boundaries [2502.07457].

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 [2101.06329] and Bayesian Faster R-CNN-based cross-domain object detection [2108.12612]. 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 [2303.03770] and graph self-training with soft posteriors \(Q(\mathbf{z}_v=k)=p_\theta(y=k\mid v)\) [2503.22744]. In dense prediction, it often means uncertainty-aware *utilization* of pseudo-labels, with hard supervision in reliable regions and soft or attenuated supervision elsewhere [2104.11596] [2505.04941] [2202.13170].

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 [2101.06329]. STRUDEL estimates pixel-wise uncertainty as MC-dropout variance over \(C=10\) stochastic passes with dropout rate \(0.2\) [2104.11596]. 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 [2209.00641]. 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 [2108.12612].

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
\[
\hat{\mathcal H}(\hat p_t)=\frac{-\sum_{c=1}^C \hat p_t^c \log_2 \hat p_t^c}{\log_2 C},
\qquad
w_{x_t}= \exp\!\big(-\hat{\mathcal H}(\hat p_t)\big)
\]
[2303.03770]. In graph self-training, the uncertainty signal is simply the softness of the posterior \(Q(\mathbf{z}_v)\), with hard pseudo-labeling only when \(\max_k Q(\mathbf{z}_v=k)>\gamma\) [2503.22744]. In building damage mapping, uncertainty is operationalized by normalized prediction entropy, with pixels satisfying \(E_n<0.3\) regarded as reliable [2505.04941]. nnFilterMatch likewise uses Shannon entropy to rank unlabeled samples for filtering inside nnU-Net training [2509.19746].

A third family uses prediction consistency. In UDASOD, uncertainty is the variance of saliency predictions under reversible augmentations such as horizontal flipping, rescaling to \(224\times224\), and FDA style swapping:
\[
Var(I_t,\tilde\theta)=\mathbb{E}\left[\left(\tilde y_t^j-\frac{1}{N}\sum_{j=1}^{N}\tilde y_t^j\right)^2\right]
\]
[2202.13170]. In domain-adaptive person re-identification, pseudo-label credibility is estimated from teacher-student consistency in a reference-center space via
\[
u_i=D_{KL}(\tilde p_i\|p_i), \qquad \omega_i=\exp(-u_i),
\]
although that method is explicitly a post-assignment reweighting framework rather than pseudo-label generation in the strict sense [2012.08733].

A fourth family uses task-specific structured uncertainty. Seq-UPS estimates uncertainty over beam-search sequence hypotheses rather than single-token confidences [2209.00641]. Robust stereo matching uses pixel-level disparity-distribution variance
\[
U^i=\sum_{\forall d^i}(d-\hat d^i)^2 \sigma(-c_d^i)
\]
and a learned area-level uncertainty network to generate sparse reliable disparity pseudo-labels [2307.16509]. Self-loop uncertainty for medical segmentation uses recurrent self-supervised encoder updates over Jigsaw transformations, then aggregates segmentation predictions with weights
\[
\omega_i = 1-\frac{l_i}{\sum_{i=1}^{Q} l_i},
\qquad
y_{sl}=\sum_{i=1}^{Q}T^{-1}_{P'_i}(S_i)\cdot norm(\omega_i)
\]
to approximate ensemble-like uncertainty at lower cost [2007.09854].

Pseudo-label representations also vary. Some methods use hard labels, such as argmax pseudo-label maps in nnFilterMatch [2509.19746] or accepted graph node labels \(\hat y_v\) after thresholding [2503.22744]. Some use soft labels, such as self-loop uncertainty maps \(y_{sl}\) [2007.09854] and saliency pseudo-labels \(\hat y_t\in[0,1]^{H\times W}\) [2202.13170]. 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 [2505.04941]. UPS additionally generalizes pseudo-labeling to include negative pseudo-labels for class absence [2101.06329].

## 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:
\[
g_c^{(i)}
=
\mathbbm{1}\!\left[u\!\left(p_c^{(i)}\right)\le \kappa_p\right]\mathbbm{1}\!\left[p_c^{(i)}\ge \tau_p\right]
+
\mathbbm{1}\!\left[u\!\left(p_c^{(i)}\right)\le \kappa_n\right]\mathbbm{1}\!\left[p_c^{(i)}\le \tau_n\right]
\]
[2101.06329]. UDTS follows the same logic in long-tailed semi-supervised learning, using adaptive uncertainty and confidence thresholds that change over training and across classes [2401.04435]. In graph learning, nodes with \(\max_k Q(\mathbf{z}_v=k)\le\gamma\) remain unlabeled, so ambiguous nodes are delayed until later EM iterations [2503.22744].

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 [2109.14651]. STRUDEL uses uncertainty-aware BCE
\[
\mathcal{L}_{\text{UBCE}}=
-\frac{1}{H\times W}\sum_{n=1}^{H\times W}(1-\sigma_n)\left[\tilde y_n\log \hat y_n +(1-\tilde y_n)\log(1-\hat y_n)\right]
\]
so highly uncertain pseudo-labeled pixels contribute less to training [2104.11596]. UDASOD uses pixel-wise weights
\[
\omega_t=\exp(-k\,Var(I_t,\tilde\theta)), \qquad k=20,
\]
while separately selecting low-uncertainty images at the sample level [2202.13170]. Source-free classification applies a smooth exponential reliability weight \(w_{x_t}=e^{-\hat{\mathcal H}(\hat p_t)}\), and reports that this soft weighting outperforms hard entropy-threshold filtering [2303.03770].

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 \(2:1\) weight ratio between the low- and high-uncertainty regions [2505.04941]. 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 [2502.07457]. 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-\(B\) sequence hypotheses [2209.00641]. Source-free classification replaces single-sample predictions with neighborhood-averaged class distributions before computing entropy [2303.03770]. UES generates pseudo-labels from an uncertainty-weighted multi-head ensemble and maps sample uncertainty to long-tailed utility weights
\[
w_i^S=\frac{1}{\frac{u_i^S}{\mathrm{MAX}(\mathcal U^S)}+1},
\]
thereby avoiding binary accept/reject decisions at the sample level [2503.09974]. 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 [2108.12612]. 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 [2109.14651]. 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 [2303.03770].

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 [2104.11596]. Self-loop uncertainty generates weighted soft pseudo-label maps by recurrently updating the encoder on a Jigsaw pretext task, then supervising only pixels with \(y_{sl}>th\) [2007.09854]. 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 [2502.07457]. 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 [2509.19746].

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
\[
P_f=\frac{P_1+De(P_2)}{2},
\]
with \(P_1,P_2\) averaged across multiple PVT-v2 backbones and flip views, after which low-uncertainty pixels satisfy \(E_n<0.3\) and receive hard supervision [2505.04941]. The damage stage mirrors this design and additionally uses a building prior to modulate features for damage pseudo-label refinement [2505.04941].

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 [2209.00641]. 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 [2307.16509].

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 [2503.22744]. UPS introduces positive and negative pseudo-labels with MC-Dropout uncertainty gating [2101.06329]. UDTS adapts uncertainty thresholds dynamically across classes and training stages for long-tailed SSL [2401.04435]. UES replaces fixed confidence thresholds with uncertainty-aware long-tailed sample weights and uncertainty-weighted ensemble heads, integrating into DualPose and FixMatch [2503.09974].

## 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 \(72.03\%\), confidence-based selection reached \(77.44\%\), 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 [2101.06329]. The same study reports that adding uncertainty on top of confidence-based selection yields a further \(5.54\%\) error reduction on CIFAR-10 with 1000 labels [2101.06329].

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 [2303.03770]. 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 [2303.03770].

In graph learning, UGST improves mean accuracy and stability: **Cora** \(83.1\pm0.8\) versus \(80.2\pm1.1\) for vanilla self-training, **Citeseer** \(73.0\pm1.0\) versus \(70.6\pm1.4\), and **Pubmed** \(80.8\pm0.9\) versus \(78.3\pm1.2\), with gains reported up to **2.5%** and lower variance across runs [2503.22744]. 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%** [2307.16509]. Threshold sweeps show that the most accurate pseudo-labels are not necessarily optimal for adaptation: for area-level uncertainty, \(t=0.1\) yields pseudo-label D1\_all \(0.30\%\) at density \(24.97\%\), but the best adaptation result is **3.64%** at \(t=0.2\), where pseudo-label D1\_all is \(0.46\%\) and density is \(45.07\%\) [2307.16509]. 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 [2505.04941]. 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 [2509.19746].

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 [2503.22744]. The source-free classification method uses neighborhood entropy as a reliability proxy rather than a formal Bayesian estimate [2303.03770]. Building damage mapping relies on normalized prediction entropy with a fixed threshold \(0.3\) and no explicit calibration step [2505.04941]. nnFilterMatch likewise depends on single-model predictive entropy without teacher models, ensembles, or calibration [2509.19746]. 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 [2503.09974].

Another limitation is **threshold and schedule sensitivity**. UPS requires \((\tau_p,\tau_n,\kappa_p,\kappa_n)\), though it reports that \(\kappa_p<0.1\) yields similar CIFAR-10 performance [2101.06329]. UGST depends on \(\gamma\) tuned over \(\{0.6,0.7,0.8,0.9\}\) [2503.22744]. Stereo adaptation uses empirically chosen thresholds such as \(t=0.9\) for \(U_{pixel}\) and \(t=0.2\) for \(U_{area}\) [2307.16509]. The remote-sensing framework fixes \(E_n<0.3\) as the low-uncertainty rule [2505.04941]. 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 [2104.11596]. BURL explicitly criticizes direct discarding of high-uncertainty regions because this may neglect valuable clinically meaningful pixels [2502.07457]. The source-free classification study shows soft exponential weighting outperforming hard entropy filtering [2303.03770]. nnFilterMatch even reverses the usual logic by removing low-uncertainty samples so training focuses on uncertain, informative cases [2509.19746]. UES goes further by assigning every sample a nonzero long-tailed weight rather than using binary acceptance [2503.09974]. 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 [2101.06329]. STRUDEL performs MC-dropout inference and iterative retraining from scratch after fine-tuning [2104.11596]. Seq-UPS combines beam search with dropout-based ensemble estimation [2209.00641]. Stereo adaptation adds an area-level uncertainty network and iterative pseudo-label regeneration [2307.16509]. 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 [2104.11596]. Building-guided damage mapping depends on the quality of stage-1 building priors [2505.04941]. Stereo pseudo-labeling depends on the source-pretrained matcher being strong enough to produce conservative reliable pixels [2307.16509]. 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.

Source: https://www.emergentmind.com/topics/uncertainty-aware-pseudo-label-generation-upg