Abstract
Deep learning models often exhibit overconfidence when faced with noisy or out-of-distribution data. We present ALUM (Adversarial Data Uncertainty Modeling), a novel framework designed to simultaneously mitigate model and data uncertainty. By leveraging latent layer statistics for compensation and mining adversarial triplets directly from training batches, ALUM achieves state-of-the-art performance in facial expression recognition and domain generalization with minimal computational overhead.
Introduction
In real-world applications, deep learning models must contend with two primary forms of uncertainty: Model (Epistemic) Uncertainty, arising from insufficient training data or parameter limitations, and Data (Aleatoric) Uncertainty, stemming from intrinsic noise like incorrect labels or sensor blur. Current methodologies often suffer from "variance diminishing," where predicted uncertainty collapses during training, rendering it ineffective for identifying difficult samples. Furthermore, most approaches focus solely on the final output layer, neglecting the rich uncertainty information present in intermediate latent layers.
The ALUM Framework
ALUM introduces a unified, model-agnostic scheme to address these challenges through three core components:
1. Latent Uncertainty Modeling and Compensation
Instead of treating intermediate features as deterministic, ALUM models each channel in a latent layer as a Gaussian distribution. By calculating channel-wise means and variances, the framework characterizes internal model uncertainty. These statistics are then used to "re-normalize" and refine feature maps, acting as a compensation mechanism that stabilizes the network against noise propagation.
2. Adversarial Triplet Mining and Mixup
Traditional adversarial training is computationally expensive, requiring multiple gradient steps to generate perturbations. ALUM bypasses this by identifying the "hardest" positive and negative examples already present within a training batch. It then performs an uncertainty-aware feature mixup using Hadamard operations (element-wise multiplication and division). This forces the model to learn from informative, difficult samples without the cost of traditional adversarial generation.
3. Self-Paced Learning Strategy
To prevent model instability during early training phases, ALUM employs a "from-easy-to-hard" strategy. The percentage of mined adversarial triplets is gradually increased as the model matures, ensuring a stable convergence path.
Experimental Results and Domain Generalization
The authors evaluated ALUM across several benchmarks, including RAF-DB for facial expression recognition and PACS for domain generalization.
Robustness to Label Noise
On the RAF-DB dataset with 30% synthetic label noise, ALUM achieved an accuracy of 84.91%, outperforming previous state-of-the-art methods by over 4%. This demonstrates the framework's ability to effectively filter and compensate for aleatoric noise.
Generalization Across Domains
In the PACS benchmark—which requires models to generalize across Photos, Art, Cartoons, and Sketches—ALUM showed significant stability. It was particularly effective in the "Sketch" domain, which is traditionally the most difficult for deep networks, exceeding the baseline performance by 4.3%.
Accuracy-Rejection Tradeoff
ALUM produces meaningful uncertainty scores. When the top 30% most uncertain samples were rejected, accuracy on the remaining test set rose to 96.46%, significantly higher than competing uncertainty-aware models.
Conclusion
ALUM provides a practical and efficient solution for modeling uncertainty in deep learning. By adding only approximately 1.5 seconds of training time per epoch on standard hardware, it offers a scalable path toward more robust AI systems. Future research will focus on establishing formal mathematical bounds for the latent compensation mechanism and testing the framework on larger-scale datasets like ImageNet.
Summary updated