[Paper Review] Generalisation in humans and deep neural networks

Robert Geirhos, Carlos R. Medina Temme+4 more|arXiv (Cornell University)|Aug 27, 2018

Infrared Target Detection Methodologies72 references239 citations

TL;DR

This study compares human object recognition robustness to twelve image degradations with three pre-trained DNNs (ResNet-152, GoogLeNet, VGG-19), showing humans generalize better to unseen distortions; training DNNs directly on distortions yields strong performance only within trained distortions and poor generalization to new ones.

ABSTRACT

We compare the robustness of humans and current convolutional deep neural networks (DNNs) on object recognition under twelve different types of image degradations. First, using three well known DNNs (ResNet-152, VGG-19, GoogLeNet) we find the human visual system to be more robust to nearly all of the tested image manipulations, and we observe progressively diverging classification error-patterns between humans and DNNs when the signal gets weaker. Secondly, we show that DNNs trained directly on distorted images consistently surpass human performance on the exact distortion types they were trained on, yet they display extremely poor generalisation abilities when tested on other distortion types. For example, training on salt-and-pepper noise does not imply robustness on uniform white noise and vice versa. Thus, changes in the noise distribution between training and testing constitutes a crucial challenge to deep learning vision systems that can be systematically addressed in a lifelong machine learning approach. Our new dataset consisting of 83K carefully measured human psychophysical trials provide a useful reference for lifelong robustness against image degradations set by the human visual system.

Motivation & Objective

• Assess how humans and current DNNs generalize to a wide range of image distortions beyond the training distribution.
• Quantify robustness differences between humans and DNNs across 12 distortion types.
• Evaluate whether training DNNs on distorted images improves cross-distortion generalization.
• Provide a carefully measured human-robot benchmark dataset for lifelong robustness against image degradations.

Proposed method

• Compare humans and three pre-trained DNNs (ResNet-152, GoogLeNet, VGG-19) on 12 distortions using a 16-class ImageNet-based 16-entry-level category task.
• Use a controlled, 200 ms presentation time with a 1/f noise mask to limit feedback in humans.
• Evaluate performance on distortions from colour changes, noise (uniform and salt-and-pepper), blur/high-pass/low-pass filters, contrast, phase noise, Eidolon distortions, and rotation.
• Train networks from scratch on distorted 16-class-ImageNet images to test distortion-specific robustness and cross-distortion generalization.
• Analyze classification accuracy and response distribution entropy to characterize error patterns and biases.

Experimental results

Research questions

• RQ1 How robust are humans versus DNNs to a variety of image degradations not seen during training?
• RQ2 Do DNNs trained on distortions generalize to other, unseen distortions?
• RQ3 Can training DNNs on distorted images improve robustness across multiple distortions, not just the trained ones?
• RQ4 What are the differences in error patterns between humans and DNNs under degraded conditions?

Key findings

• Humans are more robust than DNNs across most distortions as signal strength decreases.
• DNNs trained on distortions excel on the exact distortion they were trained on but generalize poorly to other distortions.
• Training on distortions often does not transfer robustly to unseen distortions, and may require longer training or different strategies.
• DNNs exhibit distortion-specific biases in prediction (e.g., bottle bias under strong uniform noise; dog/bird bias under phase noise).
• When trained on all distortions except one, networks reach high accuracy on those eight distortions but remain near chance for the left-out distortions (salt-and-pepper and uniform noise).
• A dedicated distortion-by-distortion training regime can close the gap on trained distortions but fails to achieve broad, cross-distortion robustness.