Modeling Human Annotation Uncertainty: A New Paradigm for Enhancing Neural Network Robustness

This article introduces an innovative study on GitHub—training neural networks using complete human label distributions (soft labels) instead of traditional hard labels. This method shows significant performance in handling blurry images, defending against adversarial attacks, and improving model generalization, providing a new paradigm for enhancing neural network robustness.

Traditional image classification training uses hard labels (single definite category), ignoring the cognitive uncertainty in human annotations. For example, a blurry image may receive different labels from multiple annotators, but traditional methods only retain a single label, leading to information loss. This makes models overconfident, lack awareness of uncertainty, and vulnerable to adversarial attacks.

This method uses soft labels that reflect human annotation distributions (e.g., if 7 out of 10 annotators label an image as cat, 2 as dog, and 1 as uncertain, the label is [0.7,0.2,0.1]). It trains the model to predict the complete probability distribution, using KL divergence or cross-entropy variants as the loss function. This method enables the model to learn to express uncertainty while acting as implicit data augmentation and regularization.

Experiments show that models trained with soft labels have a smaller drop in accuracy under adversarial attacks like FGSM and PGD compared to hard label models, without sacrificing performance on clean data; they have stronger generalization in new environments with different distributions; and they can provide reasonable probability distributions for blurry images instead of arbitrary judgments.

Three points need to be considered for application: 1. Annotation cost: Multiple annotators are needed, but a small number of distribution annotations can already improve robustness; 2. Architecture selection: Can be applied to any classification network, requiring only adjustments to the output layer and loss function; 3. Decision strategy: A confidence threshold can be set to trigger manual review, or a trade-off between precision and recall can be made.

Future exploration directions: 1. Theoretical level: Deeply understand the mathematical mechanism of how soft labels enhance robustness; 2. Methodological level: Automatic generation of soft labels; 3. Application expansion: Extend to tasks like object detection, semantic segmentation, and NLP; 4. AI ethics relevance: Improve model interpretability and fairness.

Soft label training represents a cognitive shift in machine learning—from pursuing perfect single labels to embracing labels of real distributions. This brings AI closer to the essence of human cognition, keeps it humble and prudent in high-risk scenarios, and is a key step toward building reliable AI.