Filtering Pseudo Labels via Stochastic Consistency

45 concepts35 papers2 proposals

When training deep neural networks with limited annotations, semi-supervised learning and unsupervised domain adaptation often rely on a simple idea: use the model itself to label the unlabeled data. This process, known as pseudo-labeling, creates a self-training loop where the model gradually expands its own training set.

However, this loop contains a dangerous flaw known as confirmation bias. Deep neural networks are notoriously poorly calibrated; they can be completely wrong yet highly confident. If a model assigns a 99% softmax probability to an incorrect prediction, a standard threshold filter will accept it as a ground-truth target. The model then trains on its own mistake, reinforcing its bias and causing error propagation across iterations. As argued by Rizve et al., raw softmax confidence is a deceptive metric for filtering pseudo-labels.

To break this cycle, a powerful paradigm has emerged: stochastic consistency filtering. Instead of asking how confident a network is in a single forward pass, this approach asks how stable that prediction remains when the network or the input is subjected to random perturbations.


From Confidence to Consistency

The core intuition behind stochastic consistency filtering is that true knowledge is robust, whereas errors are fragile.

Imagine a model predicting the class of an unlabeled image. In a standard network, a single forward pass yields a maximum softmax probability. To evaluate the stability of this prediction, we can introduce stochastic perturbations into the network during inference. The most common method is Monte Carlo (MC) Dropout, where random neurons are deactivated across multiple independent forward passes.

Unlabeled Input ───► [ Model + Dropout Mask 1 ] ───► Prediction 1 ┐
                   ───► [ Model + Dropout Mask 2 ] ───► Prediction 2 ┼──► Compute Variance / Entropy
                   ───► [ Model + Dropout Mask N ] ───► Prediction N ┘

If the model's output remains identical across every pass despite different active sub-networks, the prediction is highly stable. As Chen et al. demonstrated, this prediction stability under internal architectural changes serves as a far more reliable indicator of pseudo-label quality than raw confidence scores. By measuring the predictive variance across these stochastic passes, the system can filter out overconfident but incorrect labels.


Why It Works: The Geometry of Epistemic Uncertainty

To understand why consistency tracks truth better than softmax probability, we must look at the geometry of a model's decision boundaries. Softmax probability simply measures the relative distance of a sample from the decision boundary in a single, static network configuration. It does not account for epistemic uncertainty—the uncertainty stemming from a lack of training data in that region of the feature space.

As noted by Huang et al., measuring prediction variance across multiple stochastic passes explicitly quantifies this epistemic uncertainty.

  • Deep within a learned class manifold: If an unlabeled sample lies safely inside a well-trained cluster, the surrounding feature space is dense and unambiguous. Dropping out random neurons or perturbing the weights slightly will not shift the decision boundary enough to alter the classification. The variance across passes remains low.
  • Near an ambiguous or out-of-distribution boundary: If the sample is an out-of-distribution error or sits in an unlearned region, the decision boundary is highly sensitive to minor structural changes in the network. One dropout mask might classify it as Class A, while another shifts the boundary slightly, classifying it as Class B. The variance across passes spikes.

This geometric property allows the framework to decouple confidence from certainty. For example, Liu et al. showed that in adverse conditions like heavy fog, objects might yield low softmax confidence due to poor visibility. However, if their predictions remain stable across stochastic passes, they can be safely retained as pseudo-labels. Conversely, high-confidence but unstable predictions are discarded before they can pollute the training process.


Mechanisms of Perturbation

While MC Dropout is the dominant mechanism for generating predictive distributions, researchers have expanded the concept of stochastic consistency to other forms of perturbation:

1. Architectural and Weight Perturbations

Instead of dropping out neurons, models can sample entirely different weight configurations. Loh et al. replaced the final layer of a network with an efficient Bayesian layer, calculating variance across multiple stochastic weight samples to identify unreliable targets near decision boundaries.

2. Input and Augmentation Perturbations

Consistency can also be measured by altering the data itself. Yang et al. and Xu et al. evaluated the stability of model predictions under random input augmentations and perturbations, ensuring that a pseudo-label is only trusted if the model's understanding survives visual transformations. Similarly, Wang et al. measured prediction variance across masked input subsets to autonomously curate high-quality training sets from noisy quantum property data.

3. Noisy Forward Passes

In regression tasks where softmax metrics do not exist, Mehregan et al. injected noise across multiple forward passes to compute target variance, preventing incorrect spatial coordinates from degrading an indoor localization model during source-free domain adaptation.


Application Across Domains

Stochastic consistency filtering has proven highly effective across diverse, challenging machine learning domains where label noise or domain shifts are severe.

Medical Image Segmentation

Medical imaging frequently suffers from low contrast and ambiguous boundaries. Can et al. utilized MC Dropout to identify and mask out high-uncertainty pixels during scribble-supervised segmentation, preventing errors from propagating into subsequent training iterations. Yang and Li et al. applied similar voxel-level variance filtering to ensure that models only learn from high-confidence, topologically stable regions in 3D ultrasound and capillary vessel scans.

Unsupervised Domain Adaptation (UDA)

When adapting a model from a synthetic source domain to a real-world target domain, target predictions are naturally noisy. Stiefelhagen leveraged MC dropout to quantify target uncertainty, filtering out high-variance pseudo-labels to prevent them from degrading the underlying feature alignment space. Chen et al. and Khan et al. combined this stochastic filtering with model averaging to successfully adapt segmentation models without ever accessing the original source data.

Natural Language Processing

In distantly-supervised Named Entity Recognition (NER), where external knowledge bases automatically label text, the resulting data is inherently noisy. Si demonstrated that combining standard softmax scores with uncertainty estimates from multiple dropout masks effectively isolates predictions where the model is confidently wrong. In a similar vein, Wang et al. integrated MC dropout filtering into parameter-efficient self-training for language understanding, mitigating confirmation bias while updating only a fraction of the model's weights.

By shifting the selection criteria from raw, uncalibrated confidence to stochastic consistency, these frameworks allow self-training systems to safely exploit unlabeled data while remaining insulated from their own miscalibrated mistakes.

Key Foundational Work

Tri-net for Semi-Supervised Deep Learning (2018) — contributed 2 concept elements to this direction.
Learning to Segment Medical Images with Scribble-Supervision Alone (2018) — contributed 1 concept elements to this direction.
Unbiased Semi-Supervised LF-MMI Training Using Dropout (2019) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_uncertainty_aware_pseudo_label_filtering_caf8f959,
  title = {uncertainty-aware pseudo-label filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/uncertainty-aware-pseudo-label-filtering/rdr_caf8f959274249a7a998e306cb0d6a5a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_uncertainty_aware_pseudo_label_filtering_caf8f959_v1,
  title = {uncertainty-aware pseudo-label filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/uncertainty-aware-pseudo-label-filtering/rdr_caf8f959274249a7a998e306cb0d6a5a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake