ConvNets and ImageNet Beyond Accuracy: Understanding Mistakes and\n Uncovering Biases

2017/11/30 by Pierre Stock, Stock, Pierre, Moustapha Cissé +1 · 4 citations
Computer Science · #Adversarial Robustness in Machine Learning #Anomaly Detection Techniques and Applications #Artificial Intelligence (cs.AI) #Computer Vision and Pattern Recognition (cs.CV) #Computers and Society (cs.CY) #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML)

paper · pdf · doi:10.48550/arxiv.1711.11443

openalex publication_date 2017/11/30 · openalex created_date 2022/10/06 · openalex updated_date 2026/07/28

Abstract

ConvNets and Imagenet have driven the recent success of deep learning for\nimage classification. However, the marked slowdown in performance improvement\ncombined with the lack of robustness of neural networks to adversarial examples\nand their tendency to exhibit undesirable biases question the reliability of\nthese methods. This work investigates these questions from the perspective of\nthe end-user by using human subject studies and explanations. The contribution\nof this study is threefold. We first experimentally demonstrate that the\naccuracy and robustness of ConvNets measured on Imagenet are vastly\nunderestimated. Next, we show that explanations can mitigate the impact of\nmisclassified adversarial examples from the perspective of the end-user. We\nfinally introduce a novel tool for uncovering the undesirable biases learned by\na model. These contributions also show that explanations are a valuable tool\nboth for improving our understanding of ConvNets' predictions and for designing\nmore reliable models.\n

Citations

• Gradient-based learning applied to document recognition
• The Mythos of Model Interpretability
• Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word\n Embeddings
• Methods for Interpreting and Understanding Deep Neural Networks

Cited by

• Privacy Beyond Pixels: Latent Anonymization for Privacy-Preserving Video Understanding
• Women also Snowboard: Overcoming Bias in Captioning Models
• Deep Clustering for Unsupervised Learning of Visual Features
• Deep k-Nearest Neighbors: Towards Confident, Interpretable and Robust Deep Learning

Related

• Explainability and Robustness of Deep Visual Classification Models
• Is Robustness the Cost of Accuracy? -- A Comprehensive Study on the Robustness of 18 Deep Image Classification Models
• ImageNot: A contrast with ImageNet preserves model rankings
• Why Do Deep Neural Networks Still Not Recognize These Images?: A Qualitative Analysis on Failure Cases of ImageNet Classification
• From ImageNet to Image Classification: Contextualizing Progress on Benchmarks