Robust Self Training via Multi View Consensus

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2510 papers669 conceptsfrozen 2026-10-04

In semi-supervised and unsupervised learning, the "confirmation bias" problem is a persistent hurdle. When a model generates its own training targets—known as pseudo-labeling—it risks reinforcing its own mistakes. If a model is confidently wrong about a sample, it will treat that error as ground truth, leading to a feedback loop of decaying accuracy.

To break this cycle, researchers have converged on a powerful intuition: consensus. Instead of trusting the confidence of a single model, we only trust labels where multiple independent perspectives agree. This "triangulation" strategy ensures that for an error to propagate, multiple diverse views must fail in exactly the same way—a much rarer occurrence than a single model's hallucination.

The Failure of Single Model Confidence

Standard pseudo-labeling often relies on a confidence threshold (e.g., only use labels where $p > 0.95$). However, as noted by Zhongying Deng and Carola-Bibiane Schönlieb in NorMatch, these thresholds are sensitive and often poorly calibrated. A model can be highly confident yet completely incorrect, especially when facing a distribution shift.

The consensus approach replaces this internal "certainty" with external "validation." By requiring agreement between different views of the data or different model architectures, the system filters out noise that a single model might overlook.

Triangulation via Multi-View Agreement

The most intuitive form of consensus is spatial or modal. In 3D medical imaging, Yuyin Zhou and Alan Yuille demonstrated in DMPCT that requiring agreement across axial, coronal, and sagittal planes significantly reduces label noise. If a pixel is labeled "liver" in the axial view but "stomach" in the sagittal view, the prediction is discarded.

This principle extends to multi-sensor environments like autonomous driving. In DetMatch, Chenfeng Xu and colleagues use "two teachers"—one for 2D images and one for 3D LiDAR. Because cameras and LiDAR have non-overlapping failure modes (e.g., cameras struggle in low light, LiDAR struggles with reflective surfaces), their consensus is a far more robust signal than either sensor alone.

Architectural and Functional Diversity

Consensus does not always require different sensors; it can be achieved by forcing different parts of a model to "think" differently.

  • Diverse Architectures: NorMatch pairs a discriminative classifier with a generative Normalizing Flow. Because these models learn the data distribution in fundamentally different ways, their agreement is a high-fidelity indicator of label correctness.
  • Multi-Head Consistency: MingCai Chen and Chongjun Wang propose Multi-Head Co-Training (MHCT), which uses a shared backbone with multiple classification heads. By applying "strong augmentation" to different heads, they maintain the diversity needed to ensure that consensus is not just a result of the heads mimicking each other.
  • Feature vs. Probability: In DiGA, Fengyi Shen and He Wang introduce a bilateral consensus between standard probability outputs and "feature-induced" labels derived from class centroids. This ensures that a label is only accepted if it makes sense both as a classification and as a location in the feature space.

Consensus in Multi-Source Adaptation

When adapting a model to a new domain using data from multiple sources, the challenge is that some sources may be irrelevant or misleading. Naveen Venkat and R. Venkatesh Babu argue in Your Classifier can Secretly Suffice that the collective agreement of classifiers trained on different source domains serves as a natural noise filter.

Similarly, KD3A by Haozhe Feng and Fei Wu uses a consensus vote across decentralized source models to generate high-quality pseudo-labels for a target domain. This "wisdom of the sources" prevents any single biased source model from dominating the learning process.

Why It Works: The Intersection of Errors

The mathematical intuition behind consensus is the reduction of the error space. If a single model has an error rate $\epsilon$, the probability of it providing an incorrect pseudo-label is $\epsilon$. If we have two independent models with the same error rate, the probability that they both make the same mistake is significantly lower than $\epsilon$.

By requiring consensus, we effectively move the training signal from the union of the models' predictions (which contains all their individual errors) to the intersection of their predictions. As Wei Ju notes in KGNN, while this might exclude some "informative but ambiguous" samples, the resulting training set is significantly cleaner, allowing the model to build a stable foundation before tackling more complex, low-consensus regions of the data distribution.

Papers in this version

Showing 60 of 510 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 529 papers
  2. Version 1 · 2026-10-04 · 510 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_consensus_based_pseudo_labeling_4caa8623,
  title = {consensus-based pseudo-labeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/consensus-based-pseudo-labeling/rdr_4caa8623bfde4c3b99578a1937507c7b},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_robust_self_training_via_multi_view_cons_4caa8623_v1,
  title = {Robust Self Training via Multi View Consensus},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/robust-self-training-via-multi-view-consensus/rdr_4caa8623bfde4c3b99578a1937507c7b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}