Abstract
Machine learning is increasingly deployed in safety-critical domains where robustness against adversarial attacks is crucial and erroneous predictions could lead to potentially catastrophic consequences. This highlights the need for learning systems to be equipped with the means to determine a model's confidence in its prediction and the epistemic uncertainty associated with it, 'to know when a model does not know'. In this paper, we propose a novel Random-Set Convolutional Neural Network (RS-CNN) for classification which predicts belief functions rather than probability vectors over the set of classes, using the mathematics of random sets, i.e., distributions over the power set of the sample space. Based on the epistemic deep learning approach, random-set models are capable of representing the 'epistemic' uncertainty induced in machine learning by limited training sets. We estimate epistemic uncertainty by approximating the size of credal sets associated with the predicted belief functions, and experimentally demonstrate how our approach outperforms competing uncertainty-aware approaches in a classical evaluation setting. The performance of RS-CNN is best demonstrated on OOD samples where it manages to capture the true prediction while standard CNNs fail.
Introduction
Deep learning models are notoriously overconfident. Even when presented with "Out-of-Distribution" (OoD) data—samples completely unlike anything in their training set—standard neural networks often assign a high-probability score to an incorrect class. This stems from a failure to distinguish between aleatoric uncertainty (inherent noise in the data, like a blurry image) and epistemic uncertainty (the model’s lack of knowledge due to limited training data).
Existing solutions like Bayesian Neural Networks (BNNs) or Deep Ensembles attempt to address this but often face computational overhead, difficulty in scaling to large datasets like ImageNet, or reliance on sensitive prior distributions. RS-CNN introduces a "wrapper" approach that provides reliable uncertainty estimates without sacrificing speed or scalability.
Methodology: From Points to Sets
The core innovation is the Random-Set Neural Network (RS-NN) framework. Instead of a standard softmax layer outputting a probability vector for $N$ individual classes, RS-NN predicts values for a collection of sets of classes.
Belief Functions & Random Sets
Drawing from the Dempster-Shafer theory of evidence, the model assigns "mass" to subsets of classes. If a model is unsure if an image is a "3" or a "6," it can assign mass to the set {3, 6} directly, rather than being forced to split confidence between them. This allows the model to explicitly represent ignorance.
Scalability via Budgeting
Predicting every possible subset ($2^N$) is computationally prohibitive. RS-CNN employs a "budgeting" method: it clusters features from training data using Gaussian Mixture Models (GMMs) to identify natural class overlaps. It then only predicts belief values for the $N$ original classes plus the $K$ most relevant overlapping sets, making it scalable to architectures like Vision Transformers (ViT).
Training and Inference
The network uses a modified Binary Cross-Entropy (BCE) loss that encourages high belief values for any set containing the true class. For inference, the model calculates the pignistic probability (the center of mass of the predicted belief function). Uncertainty is quantified by the "width" of the credal set—the range of probabilities consistent with the evidence.
Experimental Evaluation
The authors evaluated RS-CNN across multiple domains and architectures (ResNet50, VGG16, ViT):
- Image Classification: MNIST, CIFAR-10/100, and ImageNet-1K.
- OoD Benchmarks: SVHN, Fashion-MNIST, and ImageNet-O.
- Text Classification: BBC Text vs. Emotion Detection datasets.
Key Results and Discussion
RS-CNN consistently outperformed state-of-the-art BNNs and Ensembles. On CIFAR-10, it achieved 93.53% accuracy. On ImageNet, it reached 79.92% (Top-1).
Out-of-Distribution Detection
RS-CNN showed a significant lead in detecting unfamiliar data. On the CIFAR-10 vs. Intel Image benchmark, it achieved an AUROC of 97.39, significantly higher than BNNs (~82) or standard CNNs (~87). It also achieved the lowest Expected Calibration Error (ECE), meaning its confidence scores accurately reflect its real-world performance.
Adversarial Robustness
RS-CNN is naturally resistant to adversarial attacks like FGSM and PGD. While these attacks push standard models toward a specific wrong class, RS-CNN often responds by increasing its uncertainty (widening the credal set) or predicting a set that still contains the true label.
Under a strong PGD attack on CIFAR-10, a standard CNN's accuracy dropped to near 0%, while RS-NN maintained nearly 60% accuracy.
Conclusion and Future Work
RS-CNN offers a scalable, efficient method for quantifying epistemic uncertainty. While the budgeting step requires a pre-trained feature extractor, the inference time remains nearly identical to standard CNNs. Future work will explore extending this random-set approach to regression tasks and automating the selection of focal sets ($K$).
Summary updated