ConvNets and ImageNet Beyond Accuracy: Understanding Mistakes and Uncovering Biases

By Pierre Stock, Moustapha Cissé. 2017

180 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Safety, Robustness & Interpretability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Safety, Robustness & Interpretability

Citations in the corpus

Are we done with ImageNet? (2020, background): “Several studies have revisited common computer vision benchmarks [16, 17, 18, 19] and, regarding ImageNet [1] speciﬁcally, identiﬁed various sources of bias and noise [20, 21, 22].”
Fixing the train-test resolution discrepancy (2019, background): “It has been argued that this measure is sensitive to errors in the ImageNet labels [28].”
FairFace: Face Attribute Dataset for Balanced Race, Gender, and Age (2019, background): “In computer vision, it has been shown that popular large scale image datasets such as Imagenet are biased in terms of the origin of images (45% were from the U.S.) [56] or the underlying association between scene and race [52].”
Natural Adversarial Examples (2019, background): “This is because a large share of images contain multiple classes per image [46].”
Exposing and Correcting the Gender Bias in Image Captioning Datasets and Models (2019, methodology): “The bias in datasets makes its way into the models trained on them which has been analysed by various works like [Stock and Cisse, 2018, Beigman and Klebanov, 2009, Sweeney, 2013, Flores et al., 2016, Buolamwini and Gebru, 2018, Bolukbasi et al., 2016], show how the blatant gender bias in datasets…”
Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts (2021, background): “Less speciﬁc to the Web data are the unwanted dataset biases [13, 68, 75] that are prone to ampliﬁcation by machine learning models [11, 82].”
Towards fairer datasets: filtering and balancing the distribution of the people subtree in the ImageNet hierarchy (2019, critical): “2 Stock and Cisse [71] did not explicitly analyze the dataset but demonstrate that models trained on ILSVRC exhibit misclassifications consistent with racial stereotypes.”
Large image datasets: A pyrrhic win for computer vision? (2020, background): “0 200 400 600 800 Class index [0-999] 0 500 1000 1500 2000”
In Search of Lost Domain Generalization (2020, background): “Examples of spurious correlations commonly absorbed by learning machines include racial biases [Stock and Cisse, 2018], texture statistics [Geirhos et al., 2018], and object backgrounds [Beery et al., 2018].”
A Study of Face Obfuscation in ImageNet (2021, background): “For ImageNet, researchers have examined and attempted to mitigate issues such as geographic diversity, the category vocabulary, and imbalanced representation (Shankar et al., 2017; Stock & Cisse, 2018; Dulhanty & Wong, 2019; Yang et al., 2020).”
Measuring Model Biases in the Absence of Ground Truth (2021, background): “Examples of the negative fairness impacts of machine learning models include decreased pedestrian detection accuracy on darker skin tones [29], gender stereotyping in image captioning [6], and perceived racial identities impacting unrelated labels [27].”
Characterising Bias in Compressed Models (2020, background): “Leveraging a subset of data points to understand model behaviour or to audit a dataset ﬁts into a broader literature that aims to characterize input data points as prototypes – “most typical" examples of a class – (Carlini et al., 2019; Agarwal & Hooker, 2020; Stock & Cisse, 2017; Jiang et al., 2020”
Face Recognition: Too Bias, or Not Too Bias? (2020, background): “In essence, the source varies, whether from humans [40], data or label types [35], ML models [3, 19], or evaluation protocols [34].”
Explainable Artificial Intelligence: a Systematic Review (2020, methodology): “The participants selected in the study in [348] were presented with whole images misclassified by a DNN and the visual explanations generated by LIME and MMD-critic.”
Re-labeling ImageNet: from Single to Multi-Labels, from Global to Localized Labels (2021, background): “One of the most persistent and systematic types of label error on ImageNet is the erroneous single labels [42, 39, 49, 2], referring to the cases where only one out of multiple present categories is annotated.”
The Scientific Method in the Science of Machine Learning (2019, background): “At the same time, researchers have demonstrated that trained models fail to generalize to novel scenarios (Recht et al., 2019; Zhang et al., 2019; 2018) and suggested that this decrease in performance may be due to confounders in training data (Stock & Cisse, 2018; Zhang et al., 2017).”