Balancing Quality and Quantity in Semi-Supervised Pseudo-Labeling

This direction is merged into another direction: pseudo-label confidence thresholding. Its last version stays readable and citable here.

Version 1 of 165 papers75 conceptsfrozen 2026-10-04

Semi-supervised learning (SSL) operates on a simple, seductive premise: if you have a small amount of labeled data and a mountain of unlabeled data, you can use the model’s own predictions to "label" the unlabeled pool. This process, known as pseudo-labeling, allows a student model to learn from a teacher model's guesses.

However, this creates a "Goldilocks" problem. If you trust every guess the teacher makes, the student will inevitably learn from mistakes, leading to confirmation bias—a self-reinforcing feedback loop where the model becomes increasingly confident in its own errors. To prevent this, researchers traditionally use a confidence threshold: only predictions above a certain certainty (e.g., 95%) are converted into pseudo-labels.

The Thresholding Dilemma

The central challenge in modern SSL is the fundamental trade-off between label precision and data recall. As noted by Hao Chen and colleagues in SoftMatch, rigid thresholds create a zero-sum game. A high threshold ensures "clean" labels but starves the model of information, discarding many correct but low-confidence predictions. Conversely, a low threshold provides plenty of data but introduces noise that can destabilize training.

This tension is particularly acute in complex tasks like 3D object detection or semantic segmentation. For instance, Chuandong Liu and the authors of HSSDA point out that fixed thresholds often exclude valid objects in sparse point clouds, while Siqi Fan observes in CPCL that inclusive thresholds cause segmentation models to hallucinate features in ambiguous regions.

From Heuristics to Optimization

Historically, these thresholds were hand-tuned hyperparameters, often requiring exhaustive search for every new dataset. Recent research has shifted toward treating thresholding as a dynamic optimization problem rather than a static setting.

Harit Vishwakarma and Yi Chen introduced PabLO (Rethinking Confidence Scores and Thresholds in Pseudolabeling-based SSL), which replaces heuristic tuning with an explicit "error knob." By formulating an optimization framework, practitioners can maximize the number of pseudo-labels used while strictly controlling the tolerated error rate. Similarly, Feng Wu and the team behind DAW (Exploring the Better Weighting Function for Semi-supervised Semantic Segmentation) mathematically derived that the optimal selection mechanism is a "hard step function" whose jump point is determined by the evolving distribution of correct versus incorrect predictions.

Adapting to Distribution Shifts

Thresholds are especially brittle when the unlabeled data doesn't perfectly match the labeled data. In these scenarios, a model's confidence scores become uncalibrated—it might be highly confident but totally wrong.

To combat this, Jiachen Liang and Ruibing Hou proposed Self-Supervised Feature Adaptation (SSFA) (Generalized Semi-Supervised Learning via Self-Supervised Feature Adaptation). Instead of labeling immediately, the model first uses self-supervised tasks to adapt its features to the new data distribution. This "pre-adaptation" makes the subsequent confidence scores much more reliable, allowing for cleaner pseudo-label selection even under significant domain shift.

Other approaches, like ADT-SSL (Adaptive Dual-Threshold for Semi-Supervised Learning), use class-specific adaptive thresholds. This recognizes that some classes (like "dogs") are easier for a model to learn than others (like "mops"), and therefore require different certainty requirements to maintain the same level of quality.

Beyond Binary Selection

The most recent evolution in this field moves away from the binary "keep-or-discard" logic entirely. Instead of a hard cutoff, researchers are exploring soft weighting and continuous utility.

In SoftMatch, the authors use a Gaussian weighting function to smoothly scale the contribution of each pseudo-label based on its confidence. This ensures that even lower-confidence samples contribute to the gradient, but their potential to introduce noise is mitigated by a lower weight. This philosophy mirrors Fidelity-Weighted Learning (FWL) (Learning from Samples of Variable Quality), where Arash Mehrjou and Jaap Kamps use a Bayesian teacher to estimate the reliability of each sample, effectively "down-weighting" the noise.

By moving from rigid, manual thresholds to adaptive, distribution-aware, and continuous weighting schemes, the research community is turning pseudo-labeling from a fragile heuristic into a robust engine for learning from the unlabeled world.

Papers in this version

Showing 60 of 65 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_balancing_quality_and_quantity_in_semi_s_41549d82,
  title = {Balancing Quality and Quantity in Semi-Supervised Pseudo-Labeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/balancing-quality-and-quantity-in-semi-supervised-pseudo-labeling/rdr_41549d8296054e52a065eb200ad7c35f},
  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_balancing_quality_and_quantity_in_semi_s_41549d82_v1,
  title = {Balancing Quality and Quantity in Semi-Supervised Pseudo-Labeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/balancing-quality-and-quantity-in-semi-supervised-pseudo-labeling/rdr_41549d8296054e52a065eb200ad7c35f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}