Breaking the Confirmation Bias Loop in Pseudo-Labeling

551 concepts348 papers1 proposal

Semi-supervised learning (SSL) offers an elegant promise: if you have a small amount of labeled data and a mountain of unlabeled data, a model can bootstrap its own mastery. It does this through pseudo-labeling (or self-training), where the model predicts labels for the unlabeled data, selects its most confident predictions, and treats them as ground-truth targets for its next training iteration.

But this bootstrapping process contains a dangerous flaw. If the initial labeled dataset is small or imbalanced, the model naturally learns common or "easy" classes faster than rare or "hard" ones. When it generates pseudo-labels, it will be highly confident about those easy classes and deeply uncertain about the hard ones. If left unchecked, this creates a destructive, self-reinforcing feedback loop known as confirmation bias.

The Mechanism: How Fixed Thresholds Amplify Skew

To understand why this loop occurs, consider how standard pseudo-labeling algorithms filter data. To ensure the model doesn't train on noisy, incorrect guesses, practitioners typically apply a fixed confidence threshold (e.g., only keeping pseudo-labels where the model is at least 95% confident).

While intuitive, this threshold acts as a filter that systematically favors majority classes. As noted by Jun Hyun Nam et al. in Spread Spurious Attribute: Improving Worst-group Accuracy with Spurious Attribute Estimation, fixed thresholds tend to exclude minority samples entirely because the model learns majority groups more quickly.

[Initial Imbalanced Data] ──> [Model Biased Toward Majority]
          ▲                                    │
          │                                    ▼
[Train on Skewed Labels] <── [Fixed Threshold Filters Out Minority]

This creates a "rich-get-richer" effect. For instance, in semantic segmentation, a model might easily reach a 95% confidence score for common classes like "road" or "sky" while struggling to exceed 60% for a "traffic sign." As a result, the pseudo-labeled training pool becomes entirely dominated by roads and sky. As Yen-Cheng Liu et al. observe in Unbiased Teacher for Semi-Supervised Object Detection, the model overfits to its own biased predictions, continuously strengthening its initial prejudices and progressively losing its ability to learn rare class features.

In fact, Jaehyung Kim et al. demonstrate in Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised Learning that this bias amplification can cause the inclusion of unlabeled data to hurt performance, making the model perform worse on minority classes than if it had been trained on the small labeled dataset alone.

To solve this problem, researchers have developed three core strategies to interrupt the feedback loop before it solidifies error propagation.

Strategy 1: Adaptive and Class-Specific Thresholds

If a flat confidence threshold excludes hard classes, the most direct solution is to make the threshold dynamic. Instead of forcing every class to clear the same high bar, we can lower the bar for classes the model is struggling to learn.

Yang Zou et al. pioneered this approach with Class-Balanced Self-Training (Unsupervised Domain Adaptation for Semantic Segmentation via Class-Balanced Self-training). Instead of a global threshold, the authors determine class-specific thresholds based on the model's relative confidence within each category. This ensures that a fixed proportion of pixels from every class is selected for the next training round, preventing common categories from overwhelming the dataset.

Similarly, the distribution of pseudo-labels can be monitored in real-time to gauge learning difficulty. As Chao Wang et al. show in Smoothed Adaptive Weighting for Imbalanced Semi-Supervised Learning: Improve Reliability Against Unknown Distribution Data, tracking which categories the model struggles to predict confidently allows the system to dynamically adjust the importance of the loss function, preventing the model from falling into a confirmation bias loop.

Strategy 2: Distribution Alignment

Another way to break the loop is to force the model's pseudo-label predictions to match a target distribution, preventing a single class from expanding exponentially.

When the true class distribution of the unlabeled data is known or can be estimated from the labeled set, we can treat pseudo-labeling as a constrained optimization problem. For example, the framework proposed by Jaehyung Kim et al. (Distribution Aligning Refinery of Pseudo-label for Imbalanced Semi-supervised Learning) takes the raw, biased predictions of the model and refines them so that the aggregate pseudo-label distribution matches the true class prior.

When the labeled and unlabeled data distributions do not match, more sophisticated alignment is required. Zhen Zhao et al. address this in DC-SSL: Addressing Mismatched Class Distribution in Semi-supervised Learning by using an Exponential Moving Average (EMA) model to estimate the true distribution of the unlabeled pool dynamically, rectifying the pseudo-labels before they can feed back into the training loop and entrench the model's biases.

Strategy 3: Dual-Stage Correction and Decoupling

A more fundamental intervention involves separating the generation of pseudo-labels from how they are sampled or penalized. When a model is inherently biased, simply adjusting the loss function is insufficient because the pseudo-labels themselves are already corrupted.

To address this "double whammy," Lan Li et al. introduce a two-stage correction framework in Twice Class Bias Correction for Imbalanced Semi-supervised Learning. The authors argue that effective mitigation requires de-biasing both the loss function and the label generation process simultaneously. By estimating the model's current bias, they mathematically adjust the model's raw outputs before thresholding occurs, ensuring minority classes are fairly represented in the pseudo-label pool.

By decoupling label generation from label sampling—or separating feature learning from classifier training—these methods successfully interrupt the self-reinforcing cycle. As demonstrated by Henry Peng Zou et al. in DeCrisisMB: Debiased Semi-Supervised Learning for Crisis Tweet Classification via Memory Bank, forcing the model to sample an equal number of high-confidence pseudo-labels for every class prevents it from over-training on categories it already understands well, preserving representation for rare but critical categories throughout the lifetime of training.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 345 papers
  2. Version 1 · 2026-10-04 · 239 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_pseudo_labeling_bias_626475f3,
  title = {pseudo-labeling bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/pseudo-labeling-bias/rdr_626475f3b5624eceabd0e9e93b4dc06a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_pseudo_labeling_bias_626475f3_v2,
  title = {pseudo-labeling bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/pseudo-labeling-bias/rdr_626475f3b5624eceabd0e9e93b4dc06a/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake