Distorted Representation Space Characterization Through Backpropagated\n Gradients

2019/08/26 by Gukyeong Kwon, Mohit Prabhushankar, Kwon, Gukyeong +5
Computer Science · Engineering · #Advanced Image Processing Techniques #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #FOS: Electrical engineering #Image and Signal Denoising Methods #Image and Video Processing (eess.IV) #Image and Video Quality Assessment #Infrared Target Detection Methodologies #electronic engineering #information engineering

paper · pdf · doi:10.48550/arxiv.1908.09998

openalex publication_date 2019/08/26 · openalex created_date 2022/07/19 · openalex updated_date 2026/07/28

Abstract

In this paper, we utilize weight gradients from backpropagation to\ncharacterize the representation space learned by deep learning algorithms. We\ndemonstrate the utility of such gradients in applications including perceptual\nimage quality assessment and out-of-distribution classification. The\napplications are chosen to validate the effectiveness of gradients as features\nwhen the test image distribution is distorted from the train image\ndistribution. In both applications, the proposed gradient based features\noutperform activation features. In image quality assessment, the proposed\napproach is compared with other state of the art approaches and is generally\nthe top performing method on TID 2013 and MULTI-LIVE databases in terms of\naccuracy, consistency, linearity, and monotonic behavior. Finally, we analyze\nthe effect of regularization on gradients using CURE-TSR dataset for\nout-of-distribution classification.\n

Citations

• No-Reference Image Quality Assessment in the Spatial Domain
• Visual Domain Adaptation: A survey of recent advances
• Information Content Weighting for Perceptual Image Quality Assessment
• Screenings in the rain
• Blind Image Quality Assessment: A Natural Scene Statistics Approach in the DCT Domain
• ImageNet classification with deep convolutional neural networks
• Representation Learning: A Review and New Perspectives
• Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks
• Video Anomaly Detection and Localization via Gaussian Mixture Fully Convolutional Variational Autoencoder
• Challenging Environments for Traffic Sign Detection: Reliability Assessment under Inclement Conditions

Related

• Gradients as Features for Deep Representation Learning
• Gradient Regularization Improves Accuracy of Discriminative Models
• Impact of Regularization on Calibration and Robustness: from the Representation Space Perspective
• A Learning Paradigm for Interpretable Gradients
• Scaleable input gradient regularization for adversarial robustness