Abstract
Standard deep learning classifiers often struggle with overconfidence, particularly when trained on noisy data. This paper presents a novel approach to uncertainty-aware classification by introducing the Uncertainty-Aware Negative Log-Likelihood (UANLL) loss function. By modeling classification as a multivariate regression task, the proposed method enables models to simultaneously predict class labels and estimate their own predictive uncertainty. Furthermore, the researchers leverage multi-view inference—aggregating predictions from multiple augmented versions of an input—to enhance robustness. Experimental results on CIFAR-10 with significant label noise demonstrate that this approach substantially outperforms traditional Cross-Entropy baselines and specialized state-of-the-art methods.
Introduction
In deep learning, classification models typically employ Cross-Entropy (CE) loss. While the resulting softmax outputs are frequently interpreted as confidence scores, they are often poorly calibrated and fail to represent true predictive uncertainty. This is especially problematic in the presence of "aleatoric" uncertainty, which arises from inherent noise in the data, such as mislabeled training samples.
In contrast, regression tasks often utilize heteroscedastic loss functions to predict both a mean and a variance, providing a natural measure of uncertainty. This paper bridges the gap between these two domains, proposing a framework where a classifier learns to "doubt" its own predictions, leading to more reliable decision-making in noisy environments.
Methodology
Uncertainty-Aware Negative Log-Likelihood (UANLL)
The core contribution is the UANLL loss function. Instead of treating classes as discrete categories, the authors model the classification task as a multivariate normal distribution with a "spherical" variance. The network is modified to output two components: a vector of class predictions ($h$) and a scalar uncertainty estimate ($s$).
The UANLL loss function is designed to penalize prediction errors while allowing the model to increase $s$ for difficult or noisy samples. By increasing $s$, the model can "explain away" high errors, though a penalty term prevents it from predicting high uncertainty for all inputs. This mechanism encourages the model to identify and flag inherently ambiguous samples during training.
Multi-View Inference and Aggregation
To further improve robustness during testing, the authors employ multi-view inference. Rather than relying on a single prediction, the model evaluates $n$ different augmented versions (e.g., random crops) of the same test image. These views are then aggregated using several strategies:
- MVM (Majority Vote): A simple mode-based aggregation.
- MVWCo/MVWCe: Votes are weighted by the model's internal confidence or predicted certainty.
- Hard vs. Soft Weighting: "Hard" weighting schemes utilize a threshold to completely discard views where the model's uncertainty exceeds a certain limit. The optimal thresholds and augmentation parameters are determined using Particle Swarm Optimization (PSO).
Experimental Results
The proposed methodology was evaluated on the CIFAR-10 dataset under varying levels of label noise: clean labels, 20% asymmetric noise, and 40% asymmetric noise.
Performance under Noise
The UANLL approach demonstrated exceptional resilience to label noise. On CIFAR-10 with 40% asymmetric noise, the UANLL model with hard confidence weighting (MVWCo-H) achieved an accuracy of 75.20%. This significantly exceeds the baseline CE model (~65.15%) and outperforms specialized methods like Co-teaching (~73.43%) and S2E (~75.08%). Even on clean data, the multi-view approach boosted accuracy from a baseline of ~87% to over 93%.
Impact of Multi-Views and Thresholding
Ablation studies highlighted the importance of the number of views and the optimization of hyperparameters. As shown in the figures below, accuracy increases with the number of views before plateauing around 40-50 views. The use of uncertainty as a filter (Hard Weighting) is particularly effective in noisy scenarios, as it allows the final prediction to be dominated by "clear" views while ignoring noisy or ambiguous ones.
Discussion and Conclusion
By treating classification through the lens of multivariate regression, this work provides a robust framework for uncertainty estimation. The synergy between the UANLL loss and multi-view inference allows models to effectively navigate high-noise environments.
However, some limitations remain. The current approach primarily addresses aleatoric uncertainty (data noise) rather than epistemic uncertainty (model knowledge gaps). Additionally, while the predicted variance improves performance, its direct interpretability as a probability of error requires further investigation. Future work could explore the application of these techniques to more complex domains such as object detection and out-of-distribution detection.
Summary updated