Human uncertainty makes classification more robust (1908.07086v1)

Abstract: The classification performance of deep neural networks has begun to asymptote at near-perfect levels. However, their ability to generalize outside the training set and their robustness to adversarial attacks have not. In this paper, we make progress on this problem by training with full label distributions that reflect human perceptual uncertainty. We first present a new benchmark dataset which we call CIFAR10H, containing a full distribution of human labels for each image of the CIFAR10 test set. We then show that, while contemporary classifiers fail to exhibit human-like uncertainty on their own, explicit training on our dataset closes this gap, supports improved generalization to increasingly out-of-training-distribution test datasets, and confers robustness to adversarial attacks.

• The paper demonstrates that leveraging human perceptual uncertainty through soft labels significantly enhances classifier generalization to out-of-distribution data.
• It introduces CIFAR10H, a dataset comprising 500,000 human judgments that enrich traditional hard labeling with nuanced, probabilistic information.
• The study finds that soft label training improves adversarial robustness, reducing crossentropy and boosting accuracy under FGSM and PGD attacks.

Analyzing Human Perceptual Uncertainty for Enhancing Classifier Robustness

The paper "Human uncertainty makes classification more robust" by Peterson et al. addresses a compelling need in the evolution of deep neural networks (DNNs) for image classification: improved generalization and robustness, especially in adversarial contexts. Although DNNs achieve near-perfect classification accuracy within standard benchmarks, their generalization to out-of-distribution data and resilience against adversarial attacks remain suboptimal. This paper advances the hypothesis that integrating human perceptual uncertainty into model training can enhance these aspects.

Core Contributions and Methodology

The authors introduce CIFAR10H, a novel dataset reflecting a distribution of human labels for the existing CIFAR10 test set. They harness human uncertainty about image classification—encompassing 500,000 judgments across 10,000 images—to form 'soft labels,' which differ from conventional hard labels by capturing richer distributional information. The central thesis is that training classifiers on these soft labels would align classifiers more closely with human perceptual organization and uncertainty, thus enhancing generalization and robustness.

The paper's methodological approach includes:

• Employing multiple CNN architectures trained with the CIFAR10H soft labels.
• Benchmarking these models' performances across diverse datasets, including traditional CIFAR10 and more challenging out-of-distribution sets like CINIC10 and ImageNet-Far.
• Evaluating model robustness through adversarial attack simulations, specifically using FGSM and PGD techniques.

Key Findings

The research presents salient results:

1. Generalization: Training with human-based soft labels yields better model generalization to novel datasets, with performance gains evident in increasingly out-of-distribution cases. Particularly notable is the improvement in second-best accuracy (SBA), suggesting enhanced handling of probabilistic distributions over image categories.
2. Robustness: The soft-label-trained models demonstrate increased resistance to adversarial perturbations, as indicated by reduced crossentropy and improved accuracy under attack conditions. This indicates a potential pathway for attaining adversarial robustness without explicit adversarial training.

Implications and Future Directions

The implications of this work span both practical and theoretical dimensions. Practically, leveraging human uncertainty can enhance the deployment of DNNs in real-world applications where unseen data or deliberate attacks are prevalent, such as autonomous driving systems. Theoretically, it poses intriguing questions about the nature of perceptual errors and how they can be computationally modeled to benefit artificial systems.

Looking forward, this approach prompts several avenues for future research:

• Scaling the collection and integration of human perceptual uncertainty across larger, more complex datasets.
• Further optimization of training pipelines to selectively incorporate perceptual uncertainty where it most impacts model performance.
• Investigation into unsupervised or weakly-supervised learning approaches that naturally capture and exploit data distribution nuances akin to human perception.

In summary, this paper contributes a meaningful stride in DNN research by highlighting the utility of human perceptual distribution in training for enhanced classifier robustness and generalization. It underscores the potential of blending cognitive insights with machine learning to address enduring challenges within the field.