Robust Evidential Learning Through Latent Consistency
CLEAR proposes a lightweight, task-agnostic post-hoc recalibrator for evidential deep learning that improves OOD and adversarial robustness without retraining the base predictor. Using held-out calibration data, it models group-conditioned latent geometry and measures conflict among perturbed latent views to detect unsupported evidence, selectively down-weighting overconfident predictions while preserving the original prediction. The method achieves strong gains across classification, regression, and object detection, though it trades off some in-distribution calibration and coverage for improved shift detection.
CLEAR is a pragmatic and broadly applicable post-hoc method that strengthens evidential uncertainty without altering base predictions. Its core idea—using latent-space perturbations to identify unsupported evidence—is elegant and empirically effective, yielding large gains on hard benchmarks such as ImageNet→CUB. Yet the approach is not free of limitations: it degrades in-distribution NLL and ECE compared with the base model, and in regression it can substantially reduce ID coverage. The theoretical guarantees are also modest, confined to showing that predictions and uncertainty ordering are preserved under the evidence-scaling operation.
The method’s principal strength is its preservation of evidential efficiency while improving robustness. By perturbing latent representations rather than inputs, CLEAR avoids the costly multiple forward passes of competitors such as S-EDL and C-EDL. The accompanying Proposition 1 formally ensures that classification predictions are unchanged and that uncertainty increases monotonically with latent conflict. Empirically, CLEAR achieves state-of-the-art adversarial AUROC across most classification benchmarks and successfully extends to regression and object detection without task-specific redesign.
CLEAR’s ability to flag distributional shifts is accompanied by tangible calibration costs. On California Housing, ID coverage falls from 93.01% to 57.18%, indicating that legitimate inputs may be rejected alongside OOD data.
The method also degrades NLL and ECE on clean data because it prioritises separation over conventional predictive calibration. Furthermore, the latent-conflict heuristic, while intuitively motivated, lacks formal links to adversarial robustness or distributional-shift bounds, and the regression formulation introduces an ad-hoc discretisation of the target space into quantile bins.
The experimental scope is impressive—spanning multiple datasets, corruptions, and three prediction tasks—but the comparative baseline set is largely confined to evidential methods, leaving broader UQ families (e.g., conformal, ensemble) unexamined. On difficult shifts such as ImageNet→CUB, several baselines exhibit degenerate behaviour (e.g., near-zero ID coverage), which limits the informativeness of the margins. That said, the timing analysis substantiates the efficiency claim, showing CLEAR adds only a modest overhead to the single-pass base model.
The paper documents the calibration and inference procedures in Algorithms 1 and 2 and reports hyperparameters including $V$, $\tau$, and $\delta$, with ablations suggesting reasonable stability. However, no code repository or data splits are linked from the arXiv page, severely hampering independent reproduction. Additionally, the regression variant requires partitioning the continuous target into $B$ quantile bins to define output groups, a design choice that introduces a dataset-specific tuning step and is only briefly discussed.
Reliable uncertainty quantification is essential for deploying deep learning models in high-stakes settings, where out-of-distribution and adversarial inputs can induce confident but unreliable predictions. Evidential Deep Learning provides efficient uncertainty estimates in a single forward pass, but can still assign high evidential strength to inputs that are poorly supported by the learned representation, such as adversarial inputs. We introduce CLEAR, a lightweight, task-agnostic post-hoc method that improves evidential robustness without retraining or altering the base prediction. Using held-out calibration data, CLEAR characterises the group-conditioned geometry of the model's latent space. At inference, it efficiently generates perturbation views directly in the latent space and measures their conflict relative to the calibrated geometry of the predicted group. High latent conflict indicates unsupported evidence, which CLEAR uses to selectively reduce evidential strength while retaining evidence for latent-consistent inputs. On ImageNet$\rightarrow$CUB, CLEAR improves OOD and adversarial AUROC by $+8.29$ and $+5.01$ while running 17.4$\times$ faster than competing post-hoc methods while preserving predictive performance across classification, regression, and object detection benchmarks.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.