Semi-supervised learning trains a classifier from a few labeled examples and many unlabeled ones, usually by asking the model to give the same answer on an unlabeled input and on slightly altered copies of it. The strongest alteration is adversarial: find the small nudge to the input that changes the model's output the most, then penalize that change. This is virtual adversarial training (VAT), where the model's own prediction stands in for the missing label. Its known weakness follows from that substitution. When the model is confidently wrong on an unlabeled point, VAT smooths the decision boundary around the wrong answer and locks the error in.
We propose to treat the clean input and its adversarial copy as two witnesses whose testimony is combined with evidence theory, a calculus for merging beliefs that can say "I don't know" as well as "class A." If both witnesses are confident but disagree, that disagreement is precise information: the model's confidence at this point is not supported by the input. Our training rule moves exactly that disputed belief into explicit ignorance, so the model learns to lower its confidence there instead of defending the wrong label. In short, the adversarial direction is chosen to expose conflict, and the conflict is converted into "I don't know" rather than smoothed away.
Core Claim
Replacing VAT's softmax output with a Dirichlet evidential head, and replacing its "match the clean prediction" target with a fused opinion that sends cross-view conflict to ignorance, should cut confirmation bias in the low-label regime. The Dirichlet evidential head comes from evidential deep learning (EDL): the network outputs non-negative evidence $e_k$ per class, and the total determines an ignorance mass $u = K/S$, with $S = \sum_k (e_k+1)$. The change should also produce models whose evidence is far harder to inflate adversarially, without the multiple test-time passes current defenses need. The new part is not "VAT with uncertainty filtering." It is a training target in which the adversarially found disagreement itself becomes the supervised quantity.
Concretely, write the model's opinion on input $x$ as belief masses $b_k = e_k/S$ plus ignorance $u$. For the clean view $x$ and the adversarial view $x+\delta$, with opinions $(b, u)$ and $(b', u')$, the conflict is
$$C(x,\delta) = \sum_{i \neq j} b_i\, b'_j ,$$
which is the mass that classical Dempster combination would throw away and renormalize. VAT chooses $\delta$ to maximize a divergence between outputs. We choose $\delta$, under a norm bound $\epsilon$, to maximize $C$, using one power-iteration step as VAT does so the per-step cost is unchanged. The pseudo-target is then the unnormalized fusion
$$\hat b_k = b_k b'_k + b_k u' + b'_k u, \qquad \hat u = u u' + C,$$
where agreement sharpens belief and conflict becomes ignorance. The model is trained on both views toward $(\hat b, \hat u)$ with gradients stopped through the target. Classical Dempster combination instead divides by $1 - C$. Under high conflict that division inflates a sliver of accidental overlap into near-certainty: this is Zadeh's paradox, as reexamined in Combination and Propagation of Uncertainty with Belief Functions - A Reexamination. A Framework for Comparing Uncertain Inference Systems to Probability shows the same failure can push fusion below chance. Keeping conflict unnormalized follows the open-world treatment of the transferable belief model in Default reasoning and the transferable belief model. The step that is ours is commitment to ignorance during training.
Where this sits in prior work
Three lines motivate the gap. VAT (Virtual Adversarial Training) supplies the perturbation machinery but uses a softmax output that cannot separate "evidence for several classes" from "no evidence at all." Adversarial self-training (Adversarial Self-Training Improves Robustness and Generalization for Gradual Domain Adaptation) shows that adversarial perturbations filter noisy pseudo-labels. It still relies on a scalar confidence cutoff and on gradual shift. Neither has a representation in which a confidently wrong point can be told apart from a correct one.
The closest work is C-EDL (Quantifying Adversarial Uncertainty in Evidential Deep Learning using Conflict Resolution; Robust Adversarial Quantification via Conflict-Aware Evidential Deep Learning). It shows that EDL has a "glass jaw": attacks drive evidence up and ignorance to near zero. Its fix measures conflict across five random label-preserving transforms at test time and decays evidence accordingly. It is post-hoc, uses random rather than worst-case views, and never trains the network. We take its central observation, that cross-view conflict reveals unsupported confidence, and move it into the training loss with an adversarially chosen view. This targets a single test-time pass.
On the semi-supervised evidential side, two methods use evidence on clean augmentations only. ETC-Net (An Evidential-enhanced Tri-Branch Consistency Learning Method) fuses three decoders by Dempster combination. ANEDL (Adaptive Negative Evidential Deep Learning for Open-set Semi-supervised Learning) suppresses evidence on outliers with a second head. We use one head and one extra view instead of three decoders, and our ignorance target gives an outlier-suppression signal without a separate head.
Conflict-discounting fusion constrains the design, since it shows that undiscounted combination of disagreeing sources fails. That literature includes trust networks (Navigating Conflicting Views), distance-based discounting (A Conflict-Guided Evidential Multimodal Fusion for Semantic Segmentation), and entropy-ordered weighting (Combine Conflicting Evidence Based on the Belief Entropy and IOWA Operator), which reports 87.46% versus 41.59% for plain Dempster on gear-fault data. All of these are supervised and use fixed sensors. We use the fact that the second "sensor" is constructed adversarially, which lets us avoid learned trust networks entirely.
Research Program
The build is a Wide-ResNet-style classifier with an evidential head trained on labeled data with the standard EDL loss. On unlabeled data it is trained with the conflict-seeking perturbation and the conflict-to-ignorance target above, plus a ramped weight. The analysis tracks, over training, the population of unlabeled points that are confidently mislabeled, meaning high belief on a wrong class under hidden ground truth. It asks whether our target shrinks that population faster than VAT, and whether $C$ concentrates on those points more than softmax confidence or clean-view ignorance does. A small theoretical piece is also planned. Under a local linear model, a point whose label flips within radius $\epsilon$ must incur conflict at least proportional to its clean belief on the predicted class. If so, conflict lower-bounds the overconfidence that VAT would entrench, and we will state the condition under which this holds.
Baselines
The naive reference is the same backbone trained only on the labeled pool, plus an evidential version trained the same way, so the gain from using unlabeled data at all is visible.
The first strong competitor is softmax VAT with entropy minimization, as in Virtual Adversarial Training. It is the direct ancestor and uses the same perturbation budget. We will rerun it with identical backbone, $\epsilon$, power-iteration count, and hyperparameter search budget. No reported number was found on Lacuna for VAT's CIFAR-10 or SVHN error. Beating it means lower error at 250 labels with no more passes per step.
The strongest accuracy competitor in the file is MAC-SSL (Multi-Augmentation-Based Contrastive Learning for Semi-Supervised Learning). It reports 94.96% accuracy on CIFAR-10 with 4000 labels and 58.21% on CIFAR-100 with 4 labels per class. It uses three augmentation strengths plus MixUp, so we will compare at matched wall-clock time and also report our method layered on its weak and strong augmentations. Beating it means exceeding 58.21% on CIFAR-100 at 400 labels under matched compute, or matching it at markedly lower compute.
For the open-set setting, ANEDL reports, on CIFAR-100 with 80 inlier classes, 91.1% outlier AUROC and 28.5% known-class error. We beat it only if we exceed both with one head.
For robustness of uncertainty, the comparison is C-EDL applied post-hoc to an EDL model trained semi-supervised. C-EDL reports adversarial coverage reduced to about 1.25% on CIFAR-10, and from 52.21% to 15.51% on MNIST to FashionMNIST. We count a win if our single-pass model matches its coverage while C-EDL uses five passes.
Three ablations isolate the key ingredient. The first replaces the adversarial view with a random augmentation, which tests whether worst-case search matters. The second replaces conflict-to-ignorance with normalized Dempster fusion, which tests whether committing conflict to ignorance matters. The third replaces conflict-seeking $\delta$ with VAT's KL-seeking $\delta$.
Why not just…?
Why not just use VAT with a confidence threshold? Thresholding removes uncertain points but keeps confidently wrong ones, which are precisely the points VAT entrenches. The file's evidence is indirect: ANEDL's open-set gains over a non-evidential baseline (91.1% versus 86.8% AUROC). The direct test is our confidently-mislabeled-population curve. If a tuned threshold shrinks that population as fast as our target, the evidential machinery is not earning its place.
Why not just drop points where the clean and adversarial argmax disagree, keeping softmax? This is the most serious threat. It captures the same flip signal with no evidence theory. Our claim is that dropping a point gives no gradient on it, whereas the ignorance target actively deflates evidence there and gives the model a calibrated "unknown" for open-set contamination. We run this flip-filter as a full baseline under identical compute. If it matches us on closed-set accuracy, we will say so and rest the contribution on open-set AUROC and adversarial coverage, where a discard rule produces no uncertainty output at all.
Why not just use Monte Carlo dropout or ensembles to filter pseudo-labels? ETC-Net's authors describe Monte Carlo uncertainty in semi-supervised training as computationally prohibitive. Adversarial training already multiplies passes. We will include a five-member Monte Carlo dropout filter on CIFAR-10 at 250 labels and report both accuracy and cost.
Why not just apply C-EDL at test time? It does not change what the model learned from unlabeled data, so pseudo-labels during training remain corrupted, and it costs several passes per prediction. The post-hoc baseline above measures this directly.
Proposed Experiments
The main benchmarks are CIFAR-10 at 250 and 4000 labels, CIFAR-100 at 400 labels, and SVHN at 4000 labels, with three seeds each. We report top-1 accuracy and expected calibration error. Robustness is measured as L2-PGD accuracy and adversarial coverage on the evidential predictions, following C-EDL's protocol. Open-set evaluation uses CIFAR-100 with 80 inlier classes, reporting known-class error and outlier AUROC.
The decisive checks are three. First, at 250 CIFAR-10 labels, the count of confidently mislabeled unlabeled points should fall below VAT's and below the softmax flip-filter's by mid-training. Second, the conflict $C$ should rank wrong pseudo-labels better than max softmax probability or clean-view ignorance, measured by AUROC against hidden labels. Third, the normalized-Dempster ablation should collapse toward VAT, which would confirm that sending conflict to ignorance, not the evidential head alone, drives the effect. If the random-view ablation matches the adversarial one, the adversarial-training half of the proposal fails. We would then report the method as a training-time version of C-EDL instead.
References
- Combination and Propagation of Uncertainty with Belief Functions - A Reexamination
- A Framework for Comparing Uncertain Inference Systems to Probability
- Default reasoning and the transferable belief model
- Virtual Adversarial Training: A Regularization Method for Supervised and Semi-Supervised Learning
- Adversarial Self-Training Improves Robustness and Generalization for Gradual Domain Adaptation
- Robust Adversarial Quantification via Conflict-Aware Evidential Deep Learning
- An Evidential-Enhanced Tri-Branch Consistency Learning Method for Semi-Supervised Medical Image Segmentation
- Adaptive Negative Evidential Deep Learning for Open-set Semi-supervised Learning
- Navigating Conflicting Views: Harnessing Trust for Learning
- A Conflict-Guided Evidential Multimodal Fusion for Semantic Segmentation
- Combine Conflicting Evidence Based on the Belief Entropy and IOWA Operator
- Multi-Augmentation-Based Contrastive Learning for Semi-Supervised Learning