PaperlayerView exact revision on arXiv

Source-linked AI summary

Measuring the tendency of CNNs to Learn Surface Statistical Regularities

Jason Jo, Yoshua Bengio

arXiv:1711.11561v1cs.LGstat.ML

TL;DR

The paper asks how deep CNNs can generalize well despite adversarial sensitivity, and investigates whether they exploit surface statistics rather than high-level semantics. It uses recognizability-preserving Fourier filtering to create datasets with different image statistics, finding gaps of up to 28% across test distributions and little benefit from greater depth. The results support a tendency for deep CNNs to learn surface statistical regularities, while Fourier-statistics augmentation closes the gap without necessarily producing semantic learning.

  • Problem

    The paper addresses how CNNs can generalize well despite adversarial sensitivity that casts doubt on their learning high-level abstractions.

  • Method

    The authors use tuned Fourier filtering to construct SVHN and CIFAR-10 variants that preserve human recognizability while changing surface image statistics.

  • Results

    CNNs trained on one Fourier-statistics distribution do not generalize universally across others, with gaps up to 28%, while increasing depth has little effect on closing them.

  • Takeaways & Limitations

    The findings provide empirical evidence that deep CNNs tend to rely on surface statistical regularities rather than higher-level abstractions.

  • Takeaways & Limitations

    The authors do not believe Fourier-statistics data augmentation is sufficient for learning higher-level abstractions.

Abstract

from arXiv · show

Deep CNNs are known to exhibit the following peculiarity: on the one hand they generalize extremely well to a test set, while on the other hand they are extremely sensitive to so-called adversarial perturbations. The extreme sensitivity of high performance CNNs to adversarial examples casts serious doubt that these networks are learning high level abstractions in the dataset. We are concerned with the following question: How can a deep CNN that does not learn any high level semantics of the dataset manage to generalize so well? The goal of this article is to measure the tendency of CNNs to learn surface statistical regularities of the dataset. To this end, we use Fourier filtering to construct datasets which share the exact same high level abstractions but exhibit qualitatively different surface statistical regularities. For the SVHN and CIFAR-10 datasets, we present two Fourier filtered variants: a low frequency variant and a randomly filtered variant. Each of the Fourier filtering schemes is tuned to preserve the recognizability of the objects. Our main finding is that CNNs exhibit a tendency to latch onto the Fourier image statistics of the training dataset, sometimes exhibiting up to a 28% generalization gap across the various test sets. Moreover, we observe that significantly increasing the depth of a network has a very marginal impact on closing the aforementioned generalization gap. Thus we provide quantitative evidence supporting the hypothesis that deep CNNs tend to learn surface statistical regularities in the dataset rather than higher-level abstract concepts.

1. Introduction

The paper asks how CNNs can generalize well despite adversarial sensitivity, and tests whether they rely on surface statistical regularities rather than high-level abstractions. Fourier filtering creates datasets with preserved recognizability but different image statistics, revealing substantial cross-distribution gaps.

  • CNNs achieve strong test-set generalization yet remain highly sensitive to adversarial examples, raising doubts about learned high-level abstractions.
  • The paper hypothesizes that CNNs can generalize by exploiting superficial cues shared between training and test datasets.
  • Fourier filtering produces datasets that preserve human object recognizability while exhibiting qualitatively different surface image statistics.
  • The experiments train CNNs on unfiltered, radial, or random datasets and evaluate them across all corresponding test distributions.
  • 28%: the largest reported generalization gap across Fourier-statistics test distributions.
  • 92 to 200 layers: substantially increasing CNN depth has very little effect on closing the generalization gap.
  • Fourier-statistics augmentation closes the gap, but the authors question whether it enables learning high-level semantic concepts.
  • The empirical results are presented as evidence that current deep neural networks tend to learn surface regularities instead of high-level abstractions.

2. Generalization and Surface Statistical Regularities

The paper reconciles strong CNN generalization with adversarial vulnerability by proposing that models exploit superficial image statistics shared by ordinary training and test data. This yields high performance in a narrow distributional sense without requiring explicit abstraction learning.

  • Deep CNNs can generalize extremely well while adversarial sensitivity suggests they are not truly capturing dataset abstractions.
  • Image statistics can predict object presence in natural scenes, linking superficial statistical structure to visual understanding.
  • When training and test sets share image statistics, models can generalize through superficial cues while remaining highly dependent on those statistics.
  • The main hypothesis is that deep neural networks tend to learn surface statistical regularities rather than higher-level abstractions.
  • For object recognition, these superficial cues can support high-performance generalization, but only in a narrow distributional sense.
  • Comparisons with human recognition are considered unfair because humans experience far more varied lighting, viewpoints, and occlusions.

3. Related Work

The paper positions Fourier filtering as a principled way to study surface-statistics reliance while preserving object recognizability. It relates this approach to negative-image tests, adversarial examples, robustness methods, transfer learning, and domain adaptation.

  • The authors describe Fourier filtering as the first proposed use specifically for measuring CNN reliance on surface statistical regularities.
  • Unlike negative-image tests, Fourier filtering is presented as better suited to preserving human object recognizability.
  • Negative images are more difficult for humans to recognize, making them a less controlled probe of abstraction learning.
  • Transfer learning and domain adaptation also study neural-network generalization, but the paper identifies adversarial-example research as more directly relevant.
  • Defensive distillation, contractive networks, and modified training regimes are cited as responses to adversarial vulnerability.
  • Adversarial training augments data with generated adversarial examples but is not guaranteed to resist examples produced by alternate generation methods.

4. Robustness of Object Recognition to Fourier Filtering

Fourier filtering creates low-frequency and randomly filtered image variants while preserving much of the objects’ perceptual content. The setup applies these filters to SVHN and CIFAR-10 images and evaluates whether human recognizability remains robust despite introduced artifacts.

  • Natural images concentrate much of their Fourier spectrum in low-to-medium frequencies, enabling filtering that preserves substantial perceptual content.The cited passage attributes this concentration to a power-law decay in Fourier-mode power.
  • The study compares unfiltered images with radial low-frequency filtering and uniformly random Fourier masking.Radial masking sets higher-frequency modes to zero, while random masking independently drops Fourier modes according to probability p.
  • Radial filtering uses a radius parameter, applies its mask across channels, and reconstructs images using the inverse Fourier transform.The radial variant is defined by an element-wise product between the Fourier representation and the radial mask.
  • Random filtering generates one mask and applies it consistently to all dataset elements, with the mask construction parameterized by a drop probability p.The random mask is not shared across channels.
  • For SVHN, radial radius 4.25 and random p = 0.1 preserved digit recognizability despite artifacts; for CIFAR-10, radial radius 11.0 and random p = 0.1 likewise preserved object recognizability.CIFAR-10 artifacts tended to occur in backgrounds or minimally degrade object recognition, while some SVHN random masks caused color distortions.

5. Generalization Experiments

The experiments train CNNs on unfiltered, randomly filtered, radially filtered, or fully augmented SVHN and CIFAR-10 data, then evaluate them across the corresponding test distributions. The results show substantial distribution-sensitive generalization gaps, while augmentation reduces—but may not resolve—the underlying issue.

  • Experimental setup: CNNs were trained on unfiltered, randomly filtered, or radially filtered datasets and tested across all corresponding distributions.The experiments used Preact ResNet models of depths 92 and 200, with the deepest model shown graphically.
  • SVHN experiments: SVHN models showed generalization gaps of about 6.4% to 7.89%, with the largest gaps occurring on radially filtered test data.Randomly filtered training produced the approximately 7.89% gap, while unfiltered training produced about 6.4%.
  • Cross-distribution comparison: Radially filtered training generally produced the lowest generalization gap, while no training set generalized universally well across all test sets.For SVHN, radially filtered training generalized nearly 1.5% better on the unfiltered test set than on the radially filtered test set.
  • Depth: Increasing network depth had a negligible effect on closing the observed generalization gaps across the experiments.This pattern was reported for both the SVHN and CIFAR-10 experiments.
  • CIFAR-10 experiments: CIFAR-10 models showed gaps of about 18–20% from unfiltered training and over 28% from randomly filtered training when tested on radially filtered data.These gaps were substantially larger than the analogous SVHN gap.
  • Augmentation: Fully augmented training reduced or closed the generalization gap, but the authors argue that it may provide superficial robustness rather than semantic learning.The authors compare this regularization benefit to adversarial training and question whether it enables learning higher-level semantic features.

6. Conclusion

The paper argues that deep CNNs can generalize by learning surface statistical regularities rather than high-level abstractions. Fourier filtering exposes this tendency, producing substantial generalization gaps that increased depth does not significantly close.

  • The authors hypothesize that current deep neural networks tend to learn surface statistical regularities instead of high-level abstractions.
  • Appropriately tuned Fourier filtering preserves object recognizability while changing superficial statistical regularities between clean and perturbed datasets.
  • Up to 28% test-accuracy gaps arise when CNNs trained on one class of Fourier image statistics are evaluated on qualitatively different Fourier statistics.
  • Increasing network depth does not significantly close the generalization gap, providing evidence for the paper’s main hypothesis.
  • Augmenting training with unfiltered and Fourier-filtered datasets significantly reduces the gap but is not considered sufficient for learning higher-level abstractions.

A.1. SVHN

The SVHN appendix shows additional Fourier-filtered images using random masks, including possible color deformations and two examples for each label.

  • The figure presents additional Fourier-filtered SVHN images.
  • The random mask can produce color deformations in the SVHN images.
  • Two randomly chosen images are shown for each label.

A.2. CIFAR-10

The CIFAR-10 appendix shows additional Fourier-filtered images using random masks, including possible color deformations and two examples for each label.

  • The figure presents additional Fourier-filtered CIFAR-10 images.
  • The random mask can produce color deformations in the CIFAR-10 images.
  • Two randomly chosen images are shown for each label.

B. Preact-ResNet-92 Experimental Plots

This appendix presents graphical plots for Preact-ResNet-92 and explains the Fourier-mask encoding used in Figure 9.

  • The appendix contains graphical plots for Preact-ResNet-92.
  • In Figure 9, the first image in each row corresponds to the Fourier mask in frequency space.
  • White pixels preserve Fourier modes, whereas black or colored pixels set those modes to zero.

B.1. SVHN

Figures 11 and 12 present the Preact-ResNet-92 plots for the SVHN datasets.

  • Figures 11 and 12 show the Preact-ResNet-92 plots for the SVHN datasets.

B.2. CIFAR-10

This section presents Fourier-mask visualizations and generalization plots for filtered datasets, including CIFAR-10 and SVHN figure references. The plots include unfiltered, randomly filtered, radially filtered, and fully augmented training conditions.

  • The accompanying text states that Figures 13 and 14 show Preact-ResNet-92 plots for the SVHN datasets.
  • Figure 10 visualizes Fourier masks, with white pixels preserving Fourier modes and black or colored pixels setting modes to zero.
  • Figures 11 and 12 present corresponding Preact-ResNet-92 generalization and fully augmented SVHN results.
  • Figures 13 and 14 are labeled as Preact-ResNet-92 results for CIFAR-10, including unfiltered, randomly filtered, radially filtered, and fully augmented conditions.
Loading 1711.11561v1…