Uncertainty in AI: Evaluating Deep Neural Networks on Out-of-Distribution Images

TL;DR

This paper evaluates the uncertainty and robustness of various deep neural networks on out-of-distribution and perturbed images, demonstrating ensemble methods improve classification accuracy but models remain vulnerable to perturbations.

Contribution

It introduces a comprehensive evaluation of DNNs' uncertainty on OOD data and proposes ensemble techniques to enhance classification performance.

Findings

Ensemble models outperform single models on OOD image classification.

ResNet-50 achieves the highest accuracy among tested models.

Models are highly vulnerable to image perturbations, misclassifying all perturbed images.

Abstract

As AI models are increasingly deployed in critical applications, ensuring the consistent performance of models when exposed to unusual situations such as out-of-distribution (OOD) or perturbed data, is important. Therefore, this paper investigates the uncertainty of various deep neural networks, including ResNet-50, VGG16, DenseNet121, AlexNet, and GoogleNet, when dealing with such data. Our approach includes three experiments. First, we used the pretrained models to classify OOD images generated via DALL-E to assess their performance. Second, we built an ensemble from the models' predictions using probabilistic averaging for consensus due to its advantages over plurality or majority voting. The ensemble's uncertainty was quantified using average probabilities, variance, and entropy metrics. Our results showed that while ResNet-50 was the most accurate single model for OOD images, the ensemble performed even better, correctly classifying all images. Third, we tested model robustness by adding perturbations (filters, rotations, etc.) to new epistemic images from DALL-E or real-world captures. ResNet-50 was chosen for this being the best performing model. While it classified 4 out of 5 unperturbed images correctly, it misclassified all of them post-perturbation, indicating a significant vulnerability. These misclassifications, which are clear to human observers, highlight AI models' limitations. Using saliency maps, we identified regions of the images that the model considered important for their decisions.