High-Frequency Component Helps Explain the Generalization of Convolutional Neural Networks

Wang, Haohan; Wu, Xindi; Huang, Zeyi; Xing, Eric P.

doi:10.1109/cvpr42600.2020.00871

Cited by 388 publications

(250 citation statements)

References 17 publications

Supporting

Mentioning

208

Contrasting

Order By: Relevance

“…For example, many machine-vision systems are intolerant to image distortions: If CNNs are trained on clean images but tested on noisy images, they perform far below humans at test (73). But here too, if machines were burdened with humanlike visual acuity and so could barely represent the high-frequency features in the training set (i.e., the features most distorted by this sort of noise), they may be less sensitive to the patterns that later mislead them (74). Indeed, recent work finds that giving CNNs a humanlike fovea (75) or a hidden layer simulating V1 (76)…”

Section: Limit Machines Like Humansmentioning

confidence: 99%

“…97)-which, while literally true, may not explain why such different representations arise in the first place. Indeed, it is telling that much progress in understanding such errors has come from "behavioral" studies (in humans and machines) (69,70,74,80). A mixed strategy is likely best, with behavioral comparisons as essential components.…”

Section: What Species-fair Comparisons Showmentioning

confidence: 99%

See 1 more Smart Citation

Performance vs. competence in human–machine comparisons

Firestone

2020

Proc. Natl. Acad. Sci. U.S.A.

123

View full text Add to dashboard Cite

Does the human mind resemble the machines that can behave like it? Biologically inspired machine-learning systems approach “human-level” accuracy in an astounding variety of domains, and even predict human brain activity—raising the exciting possibility that such systems represent the world like we do. However, even seemingly intelligent machines fail in strange and “unhumanlike” ways, threatening their status as models of our minds. How can we know when human–machine behavioral differences reflect deep disparities in their underlying capacities, vs. when such failures are only superficial or peripheral? This article draws on a foundational insight from cognitive science—the distinction between performance and competence—to encourage “species-fair” comparisons between humans and machines. The performance/competence distinction urges us to consider whether the failure of a system to behave as ideally hypothesized, or the failure of one creature to behave like another, arises not because the system lacks the relevant knowledge or internal capacities (“competence”), but instead because of superficial constraints on demonstrating that knowledge (“performance”). I argue that this distinction has been neglected by research comparing human and machine behavior, and that it should be essential to any such comparison. Focusing on the domain of image classification, I identify three factors contributing to the species-fairness of human–machine comparisons, extracted from recent work that equates such constraints. Species-fair comparisons level the playing field between natural and artificial intelligence, so that we can separate more superficial differences from those that may be deep and enduring.

show abstract

Section: Limit Machines Like Humansmentioning

confidence: 99%

Section: What Species-fair Comparisons Showmentioning

confidence: 99%

Performance vs. competence in human–machine comparisons

Firestone

2020

Proc. Natl. Acad. Sci. U.S.A.

123

View full text Add to dashboard Cite

show abstract

“…Ilyas et.al [6], on the other hand, claimed that among the feature patterns that can be used to train generalizable models, some are susceptible to attacks and others are not ; therefore, training robust models would require encoding human priors during training, which may not be known. Meanwhile, Wang et.al [7] found that highfrequency components of the inputs (images in this particular study) were targeted by adversarial perturbations and suggested smoothing the convolutional filters to reduce the influence of high-frequency features on the model's output. While there are some difference, a common thread in all of the aforementioned studies is that reducing the model's reliance on superfluous features can make it more robust to adversarial perturbations.…”

Section: Introductionmentioning

confidence: 85%

“…Recent studies have found that deep learning models are made vulnerable to adversarial attacks because their decision function relies on spurious features, which the adversary can perturb to induce misclassifications [5,6,7]. Building upon this line of work, in this paper we develop an approach to identify and remove spurious features from the network.…”

Section: Adversarial Attacks and Defensesmentioning

confidence: 99%

Towards Adversarial Robustness Via Compact Feature Representations

Shah

Olivier

Raj

2021

ICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

View full text Add to dashboard Cite

Deep Neural Networks (DNNs), while providing state-of-the-art performance in a wide variety of tasks, have been shown to be vulnerable to adversarial attacks. Recent studies have posited that this vulnerability arises because DNNs operate over a grossly overspecified input space with very sparse human supervision due to which they tend to learn spurious features that humans would ignore. These spurious features provide an attack vector for the adversary because perturbing these features would not alter the human's decision but may alter the model's prediction. In this paper we explore hypothesis that reducing the size of the model's feature representation while maintaining its generalizability would discard spurious features while retaining perceptually relevant ones. We find that after the size of the feature representation has been reduced the models exhibit increased adversarial robustness, while suffering only a minimal loss in accuracy. In addition to being more robust, models with compact feature representations have the benefit of being more resource efficient.

show abstract

“…As (3) and ( 4) can be used to estimate how much of highfrequency image content is partially suppressed but remains preserved by an image filtering method, it would be interesting to establish a link with the ability of the method to protect deep learning schemes against adversarial attacks which often target high-frequency image components invisible to the human eye [20,21].…”

Section: Conclusion and Directions For Future Workmentioning

confidence: 99%

Two iterative methods for reverse image filtering

Belyaev

Fayolle

2021

SIViP

View full text Add to dashboard Cite

We consider the problem of recovering an original image $${\varvec{x}}$$ x from its filtered version $${\varvec{y}}={\varvec{f}}({\varvec{x}})$$ y = f ( x ) , assuming that the internal structure of the filter $${\varvec{f}}(\cdot )$$ f ( · ) is unknown to us (i.e., we can only query the filter as a black-box and, for example, cannot invert it). We present two new iterative methods to attack the problem, analyze, and evaluate them on various smoothing and edge-preserving image filters.

show abstract

High-Frequency Component Helps Explain the Generalization of Convolutional Neural Networks

Cited by 388 publications

References 17 publications

Performance vs. competence in human–machine comparisons

Performance vs. competence in human–machine comparisons

Towards Adversarial Robustness Via Compact Feature Representations

Two iterative methods for reverse image filtering

Contact Info

Product

Resources

About