On instabilities of deep learning in image reconstruction and the potential costs of AI
О неустойчивости глубокого обучения при реконструкции изображений и потенциальных издержках искусственного интеллекта
2020-05-11
SCID: 54.1/25vesw73
Discuss with AI
adversarial perturbationsdeep learningimage reconstructionreconstruction instabilitysampling-domain perturbations
Figures from the paper
Abstract (AI)
Deep learning, due to its unprecedented success in tasks such as image classification, has emerged as a new tool in image reconstruction with potential to change the field. In this paper, we demonstrate a crucial phenomenon: Deep learning typically yields unstable methods for image reconstruction. The instabilities usually occur in several forms: 1) Certain tiny, almost undetectable perturbations, both in the image and sampling domain, may result in severe artefacts in the reconstruction; 2) a small structural change, for example, a tumor, may not be captured in the reconstructed image; and 3) (a counterintuitive type of instability) more samples may yield poorer performance. Our stability test with algorithms and easy-to-use software detects the instability phenomena. The test is aimed at researchers, to test their networks for instabilities, and for government agencies, such as the Food and Drug Administration (FDA), to secure safe use of deep learning methods.
Key Findings
1
Counterintuitively, increasing the number of samples can worsen reconstruction performance.
2
Deep learning methods for image reconstruction are typically unstable, despite their success in image classification.
3
Deep reconstruction networks may fail to capture small structural changes, such as tumors, in reconstructed images.
4
The proposed stability test and accompanying software detect these instabilities for researchers and support safety evaluation by agencies such as the FDA.
5
Tiny, nearly imperceptible perturbations in image or sampling domains can produce severe reconstruction artefacts.
Research Object
deep learning methods for image reconstruction
Research Subject
instability phenomena and associated reconstruction failures under image and sampling perturbations, structural changes, and increased sample numbers
Publication Details
Publication Date
2020-05-11
Journal
Publisher
ISSN
Cited by
792
Open access PDF
Access Type
Author Information
Download PDF
Subscribe to digest
References available in scid.ai5
ImageNet classification with deep convolutional neural networks2017
Very Deep Convolutional Networks for Large-Scale Image Recognition2014
Adam: A Method for Stochastic Optimization2014
Intriguing properties of neural networks2013
DeepFool: A Simple and Accurate Method to Fool Deep Neural Networks2016