Discrepancy-Confusion Uncertainty in DECERN

• Discrepancy-confusion uncertainty is a composite measure designed to assess informativeness in active fine-grained image classification by evaluating shifts in predictive distributions under local feature perturbations.
• It quantifies model sensitivity by calculating discrepancy via cross-entropy shifts and confusion via increased prediction entropy on locally fused features.
• The method integrates uncertainty-weighted clustering and calibration diversity to select informative, non-redundant samples for annotation in active learning pipelines.

Discrepancy-confusion uncertainty is a composite informativeness measure introduced for active fine-grained image classification in DECERN. It is designed to identify unlabeled fine-grained images whose feature representations are both directionally ambiguous with respect to any one category and structurally unstable under small, semantically-informed perturbations. In that formulation, discrepancy uncertainty measures how much the model’s predicted distribution on a perturbed feature shifts away from its original prediction, whereas confusion uncertainty measures how high the entropy of the model’s prediction becomes on the perturbed feature. The resulting score is used to form uncertainty candidates, after which uncertainty-weighted clustering and calibration diversity are applied to select annotation queries (Jin et al., 29 Sep 2025).

1. Definition and motivation

In fine-grained classification, many subcategories share very similar visual semantics, and traditional uncertainty metrics such as margin, entropy, and confidence often fail to distinguish samples that sit on subtle decision boundaries or whose internal feature structure is fragile. Discrepancy-confusion uncertainty was introduced to address that failure mode by explicitly probing category directionality and structural stability through local feature-fusion perturbations (Jin et al., 29 Sep 2025).

The measure has two components. Discrepancy uncertainty quantifies how much the model’s predicted distribution on the perturbed feature shifts away from the original prediction, implemented as a cross-entropy term between fused and original distributions. Confusion uncertainty quantifies how high the entropy of the model’s prediction becomes on the fused feature, indicating loss of directional signal. The combination is intended to produce a richer, two-dimensional view of informativeness, targeting samples that are both hard to steer toward any one class and whose local feature manifold is unstable (Jin et al., 29 Sep 2025).

This construction is specific to active learning in fine-grained settings. It does not begin from a generic pool-based uncertainty score computed directly on the original image; instead, it asks how the classifier behaves under local, class-conditioned feature perturbations that are derived from the current labeled set.

2. Mathematical construction in DECERN

Let the unlabeled pool be $\mathcal D^u=\{x_i^u\}$ and the labeled pool be $\mathcal D^l=\{(x_i^l,y_i^l)\}$. For an unlabeled sample $x^u$, DECERN defines a feature encoder $\mathcal F_z(\cdot)$ and classifier $\mathcal F_c(\cdot)$, with

$$z^u=\mathcal F_z(x^u),\qquad p^u=\mathcal F_c(z^u).$$

For class $j$, the class anchor is the average feature over labeled examples of that class,

$$z_j^a=\frac{\sum_{(x,y)\in\mathcal D^l}\mathbf 1\{y=j\}\cdot \mathcal F_z(x)}{\sum \mathbf 1\{y=j\}},$$

and the corresponding averaged class prediction is

$$p_j^a=\frac{\sum \mathbf 1\{y=j\}\cdot \mathcal F_c(\mathcal F_z(x))}{\sum \mathbf 1\{y=j\}}.$$

A binary mask $M$ is computed from the top-$\mathcal D^l=\{(x_i^l,y_i^l)\}$0 positions of the gradient $\mathcal D^l=\{(x_i^l,y_i^l)\}$1, isolating the most salient feature dimensions. The fusion strength for class $\mathcal D^l=\{(x_i^l,y_i^l)\}$2 is set as $\mathcal D^l=\{(x_i^l,y_i^l)\}$3. Local feature fusion is then defined by

$\mathcal D^l=\{(x_i^l,y_i^l)\}$4

This operator perturbs only the salient feature positions, nudging the unlabeled feature toward the class anchor in a confidence-weighted manner (Jin et al., 29 Sep 2025).

From this construction, three prediction distributions are considered for class $\mathcal D^l=\{(x_i^l,y_i^l)\}$5:

$\mathcal D^l=\{(x_i^l,y_i^l)\}$6

$\mathcal D^l=\{(x_i^l,y_i^l)\}$7

$\mathcal D^l=\{(x_i^l,y_i^l)\}$8

where the fusion ratio is

$\mathcal D^l=\{(x_i^l,y_i^l)\}$9

The category-level score is built from an entropy term and a cross-entropy term,

$x^u$0

where $x^u$1 is one of $x^u$2. A category weighting

$x^u$3

measures how far the unlabeled feature lies from the class-$x^u$4 anchor in cosine space. The category-level discrepancy-confusion score is then

$x^u$5

Finally, the instance-level score averages over all $x^u$6 classes and the three fusion modes:

$x^u$7

In this formulation, discrepancy-confusion uncertainty is therefore not a single entropy statistic. It is an aggregation over class anchors, saliency-restricted perturbations, and multiple fused distributions (Jin et al., 29 Sep 2025).

3. Candidate formation, uncertainty-weighted clustering, and calibration diversity

DECERN integrates discrepancy-confusion uncertainty into a multi-stage active learning pipeline. The first stage computes $x^u$8 for every unlabeled sample and applies dynamic thresholding. If $x^u$9 and $\mathcal F_z(\cdot)$0 are the mean and standard deviation of $\mathcal F_z(\cdot)$1 over the current unlabeled pool, the threshold is

$\mathcal F_z(\cdot)$2

with $\mathcal F_z(\cdot)$3 driven by the skewness of $\mathcal F_z(\cdot)$4. All samples with $\mathcal F_z(\cdot)$5 become uncertainty candidates (Jin et al., 29 Sep 2025).

The second stage performs uncertainty-weighted clustering. K-means is run on the candidate features $\mathcal F_z(\cdot)$6 to form $\mathcal F_z(\cdot)$7 clusters, while weighting each point by its discrepancy-confusion score. The stated purpose is to pull centroids toward high-uncertainty regions. Let the centroid of cluster $\mathcal F_z(\cdot)$8 be denoted $\mathcal F_z(\cdot)$9.

The third stage applies calibration diversity sampling. For each cluster $\mathcal F_c(\cdot)$0, one sample is selected by balancing local representativeness against global diversity:

$\mathcal F_c(\cdot)$1

where $\mathcal F_c(\cdot)$2 balances the two objectives. Local representativeness is defined as proximity to the cluster centroid, and global diversity is defined as distance from all class anchors. Selected samples are queried from the oracle, added to $\mathcal F_c(\cdot)$3, and the network is retrained in the next active learning cycle (Jin et al., 29 Sep 2025).

Within DECERN, discrepancy-confusion uncertainty thus plays an upstream filtering role rather than serving as the sole acquisition function. It defines the candidate set that is subsequently diversified by clustering and calibrated against the anchor geometry of the labeled set.

4. Relation to other discrepancy and confusion formulations

The terms discrepancy and confusion appear in several technically distinct literatures, and discrepancy-confusion uncertainty occupies a specific position among them.

In flexible visual recognition, confusion and ignorance are separated under Subjective Logic. There, confusion is defined as conflicting evidence, implemented as total mass on all non-singleton subsets, while ignorance is the absence of evidence, implemented as mass on the empty set. By contrast, discrepancy-confusion uncertainty in DECERN does not operate with hyper-opinions, non-singleton masses, or ignorance mass; its confusion term is the entropy of the classifier output on locally fused features (Fan et al., 2023).

In diffusion MRI, NUQ uses the maximum mean discrepancy metric to compare samples drawn from posterior distributions of microstructure measures. The resulting score is a pooled quality score that captures noise-induced uncertainty by comparing entire posterior distributions. This is again different from DECERN: the discrepancy in NUQ is a kernel two-sample discrepancy between posterior samples, whereas the discrepancy term in discrepancy-confusion uncertainty is a cross-entropy shift between original and fused predictive distributions (Fadnavis et al., 2022).

In open-set medical image segmentation, ASFOSDA combines Class-aware Decomposed Uncertainty and Class-agnostic Prototype Discrepancy. There, uncertainty is decomposed into aleatoric and epistemic components, and discrepancy is defined through cross-domain and self-domain cosine-based prototype differences. This differs from DECERN’s category-wise local feature fusion, but both methods combine uncertainty estimation with a diversity mechanism rather than relying on uncertainty alone (Yang et al., 7 Jun 2026).

In mechanistic electrophysiology, model discrepancy denotes the model–reality gap and is represented either explicitly as a discrepancy function $\mathcal F_c(\cdot)$4 added to the simulator output or empirically through the variability of predictions from parameter sets trained on different experimental protocols. Those usages concern misspecification in scientific simulators, not class-boundary ambiguity in discriminative representation learning (Lei et al., 2020, Shuttleworth et al., 2023).

In quantum measurement theory, discrepancy-based error definitions compare actual and ideal values, while an operational approach defines error and disturbance through the distinguishability of devices. That literature uses “confusion” in an operational sense tied to distinguishability tests rather than entropy under feature perturbation (Renes et al., 2016).

Taken together, these formulations show that discrepancy-confusion uncertainty is a domain-specific construct: it is neither synonymous with model discrepancy, nor with kernel discrepancy, nor with evidential confusion.

5. Empirical behavior in active fine-grained classification

DECERN was evaluated on 7 fine-grained image datasets across 26 distinct experimental settings, using ResNet50 and ViT backbones and budgets $\mathcal F_c(\cdot)$5 with $\mathcal F_c(\cdot)$6. Table 1 reports that DECERN outperforms 8 state-of-the-art active learning methods, and in every setting it is best or second-best. Figure 1 shows accuracy over 8 active learning cycles on Caltech101, StanfordDogs, and BronzeDing, with DECERN’s curve consistently above competitors (Jin et al., 29 Sep 2025).

The ablation study directly isolates the role of discrepancy-confusion uncertainty. Table 3 shows that removing either the discrepancy term $\mathcal F_c(\cdot)$7 or the confusion term $\mathcal F_c(\cdot)$8 reduces accuracy by 1–2 points. The same table also reports that dropping the uncertainty-weighted clustering or the calibration diversity component harms performance. The t-SNE visualizations in Figure 2 show that DECERN’s selected samples cover decision boundary regions more uniformly and avoid redundant clusters seen in other active learning methods (Jin et al., 29 Sep 2025).

Hyperparameter studies in Figure 3 report stable performance for moderate fusion ratio $\mathcal F_c(\cdot)$9, dynamic thresholding $$z^u=\mathcal F_z(x^u),\qquad p^u=\mathcal F_c(z^u).$$0, and diversity balance $$z^u=\mathcal F_z(x^u),\qquad p^u=\mathcal F_c(z^u).$$1. Within the reported evidence, discrepancy-confusion uncertainty is therefore not presented as an isolated scoring heuristic; it is validated as part of a selection pipeline in which uncertainty, clustering, and diversity calibration jointly determine query quality (Jin et al., 29 Sep 2025).

6. Interpretation, scope, and common misconceptions

A common simplification is to equate discrepancy-confusion uncertainty with ordinary predictive entropy. That is not the construction used in DECERN. Traditional entropy and margin operate in the original feature space, whereas discrepancy-confusion uncertainty explicitly simulates semantically-informed local perturbations through class anchors, a salient-dimension mask, and confidence-weighted fusion. The resulting score tests whether the model’s decision flips drastically and whether the prediction becomes highly uncertain under that perturbation (Jin et al., 29 Sep 2025).

A second simplification is to treat the method as a generic diversity sampler. In DECERN, diversity enters after uncertainty candidate formation. Uncertainty-weighted clustering is applied only to samples with $$z^u=\mathcal F_z(x^u),\qquad p^u=\mathcal F_c(z^u).$$2, and calibration diversity then trades off local representativeness against global diversity relative to class anchors. This suggests that the method is intended to select samples that are simultaneously uncertain, non-redundant, and anchor-distant, rather than globally diverse in isolation (Jin et al., 29 Sep 2025).

A third potential confusion concerns the word confusion itself. In DECERN, confusion uncertainty is the entropy of the classifier’s output on a locally fused feature. In evidential flexible recognition, confusion is total mass on non-singleton subsets and is explicitly separated from ignorance, the mass on the empty set. The two notions address different decision problems: one probes local manifold fragility in fine-grained active learning, and the other supports flexible recognition through multi-class prediction and rejection (Fan et al., 2023).

The scope of discrepancy-confusion uncertainty is therefore narrow but technically precise. It is a sample-selection criterion for active fine-grained classification, built from local feature fusion with class anchors, category-wise discrepancy and confusion terms, dynamic thresholding, and downstream diversity calibration. A plausible implication is that it is especially suited to regimes in which subtle inter-class differences, rather than coarse class separation or out-of-distribution rejection, dominate the informativeness of unlabeled data.