Improving the robustness of ImageNet classifiers using elements of human visual cognition

A. Emin Orhan, Brenden M. LakearXiv 2019 · OpenReview

TL;DR

This paper explores how to make ImageNet classifiers more robust to adversarial and natural image perturbations by incorporating two human vision principles: episodic memory and a shape bias. They found that while episodic memory helps against small adversarial attacks, a shape bias is crucial for robustness against natural corruptions, and combining both offers synergistic improvements.

What Problem Does This Paper Address?

Deep neural networks (DNNs) are state-of-the-art for computer vision tasks and even serve as models for human visual systems. However, they suffer from several critical deficiencies compared to human vision: extreme sensitivity to imperceptible adversarial perturbations, greater sensitivity to natural image distortions, a strong reliance on local texture over global shape, and a lack of an "episodic memory" component that allows humans to recognize objects based on familiarity or novelty. This paper aims to address these robustness issues by integrating elements of human visual cognition into DNNs, specifically focusing on adversarial and natural perturbations at the ImageNet scale.

Methods

The core of the proposed method involves augmenting standard ImageNet-trained models (specifically ResNet-50) with an explicit episodic cache memory. This cache stores low-dimensional embeddings (keys) and corresponding class labels (values) for all training images. When a new image is presented, its embedding is compared to the cache keys, and a prediction is made based on a weighted combination of the cached values.

The authors also explore the impact of using shape-based features derived from a model trained to prioritize global shape over local texture (Shape-ResNet-50) as the embeddings for the cache.

Several design choices for the episodic memory are investigated:

  • Retrieval Method: Comparing a "continuous cache" (using all items weighted by similarity) with a k-nearest neighbor (k-NN) approach.
  • Cache Size: Reducing the number of items in the cache through clustering.
  • Dimensionality: Reducing the dimensionality of the embeddings.

Datasets

  • ImageNet: Used for training and evaluating the models. The paper mentions using approximately 1.28 million training images and 50,000 validation images across 1000 classes.
  • ImageNet-C: A benchmark dataset for measuring robustness against natural perturbations. It contains 15 different types of corruptions (noise, blur, weather, digital) applied at 5 severity levels to the ImageNet validation set.

Key Insights and Ideas

  • Episodic Memory for Adversarial Robustness (but not natural): An explicit episodic memory cache significantly improves robustness against small-norm adversarial perturbations (especially in gray-box and black-box settings) but offers no improvement against larger, natural perturbations. This suggests caching acts as a regularizer for specific types of attacks.
  • Shape Bias is Key for Natural Robustness: Models that learn to prioritize global shape over local texture (like Shape-ResNet-50) are inherently more robust to natural image corruptions. This aligns with human visual perception.
  • Synergistic Improvement: Combining an episodic memory with shape-based features leads to a synergistic improvement in robustness, outperforming either approach alone. This suggests that learning robust features first and then augmenting with memory is a powerful strategy.
  • Compact Caches: When making caches more compact for scalability, clustering items (reducing cache size) is more effective than reducing embedding dimensionality, preserving accuracy better.
  • White-box Attacks Remain Challenging: Even with these improvements, white-box adversarial attacks remain difficult to defend against, suggesting that test-time interventions like caching might not be sufficient for the strongest attack scenarios.

Key Results

  • Adversarial Robustness: Cache models using shape-based features achieved significantly higher adversarial accuracy in gray-box settings compared to texture-based features. For example, with a perturbation size of $\epsilon = 0.06$, the cache with shape features achieved 0.336 accuracy, while the texture-based cache achieved 0.170.
  • Natural Robustness: On ImageNet-C, the baseline ResNet-50 had a mean corruption error (mCE) of 0.764. Shape-ResNet-50 reduced this to 0.735. Caching alone did not improve mCE. However, caching with shape-based features maintained this improved mCE of 0.735.
  • Cache Design Trade-offs: Reducing cache dimensionality by 8x led to a clean accuracy of 49.8%, while reducing cache size by 8x resulted in a clean accuracy of 61.2%, highlighting the preference for size reduction over dimensionality reduction for compactness.
  • White-box Performance: The white-box adversarial setting showed limited improvement, with accuracies often remaining very low (e.g., 0.000 or 0.001) for most cache configurations, suggesting this attack is particularly difficult to defend against.

Limitations and Future Work

  • White-box Adversarial Robustness: The paper acknowledges that caching does not significantly improve robustness against white-box adversarial attacks. This is attributed to the inherent difficulty of defending against such strong attacks with test-time interventions.
  • Scalability: While the paper discusses making caches more compact, scaling to extremely large datasets (beyond ImageNet) would require further investigation.
  • Future Directions: The authors are interested in applying these naturally and adversarially robust features to few-shot recognition tasks and in modeling neural and behavioral data from humans and monkeys. They also suggest exploring architectural inductive biases that might inherently lead to better shape integration.

Summary updated

Abstract

We investigate the robustness properties of image recognition models equipped with two features inspired by human vision, an explicit episodic memory and a shape bias, at the ImageNet scale. As reported in previous work, we show that an explicit episodic memory improves the robustness of image recognition models against small-norm adversarial perturbations under some threat models. It does not, however, improve the robustness against more natural, and typically larger, perturbations. Learning more robust features during training appears to be necessary for robustness in this second sense. We show that features derived from a model that was encouraged to learn global, shape-based representations (Geirhos et al., 2019) do not only improve the robustness against natural perturbations, but when used in conjunction with an episodic memory, they also provide additional robustness against adversarial perturbations. Finally, we address three important design choices for the episodic memory: memory size, dimensionality of the memories and the retrieval method. We show that to make the episodic memory more compact, it is preferable to reduce the number of memories by clustering them, instead of reducing their dimensionality.
BibTeX
@misc{orhan2019improving,
  author = {A. Emin Orhan and Brenden M. Lake},
  title = {{Improving the robustness of ImageNet classifiers using elements of human visual cognition}},
  year = {2019},
  eprint = {1906.08416},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1906.08416},
}
Figure