Generalisation in humans and deep neural networks

By Robert Geirhos, Carlos R. Medina Temme, Jonas Rauber, Heiko H. Schütt, Matthias Bethge, Felix Wichmann. 2018

685 citations recorded by the source corpus.

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.

Subfields: Scaling & Learning Theory

How researchers see this paper

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

In Scaling & Learning Theory

Citations in the corpus

Shortcut learning in deep neural networks (2020, background): “While adversarial attacks test model performance on model-dependent worst-case noise, a related line of research focuses on model-independent noise like image corruptions [51, 52].”
Measuring Robustness to Natural Distribution Shifts in Image Classification (2020, methodology): “Our testbed consists of 109 model predictions and is 100 times larger than prior work [27, 33, 47, 68].”
The Many Faces of Robustness: A Critical Analysis of Out-of-Distribution Generalization (2020, background): “For instance, Texture Bias was initially investigated with synthetic distortions (Geirhos et al., 2018), which conflicts with the Synthetic 6 =⇒ Natural hypothesis.”
Do Image Classifiers Generalize Across Time? (2019, background): “Recent work has proposed less adversarial image modifications such as small rotations & translations [6, 1, 7, 17], hue and color changes [14], image stylization [9] and synthetic image corruptions such as Gaussian blur and JPEG compression [13, 10].”
Do ImageNet Classifiers Generalize to ImageNet? (2019, neutral): “, [17, 23].”
ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness (2018, methodology): “In combination with previous work which showed that changing other major object dimensions such as colour (Geirhos et al., 2018) and object size relative to the context (Eckstein et al.”
AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (2019, background): “Mixing augmentations allows us to generate diverse transformations, which are important for inducing robustness, as a common failure mode of deep models in the arena of corruption robustness is the memorization of ﬁxed augmentations (Vasiljevic et al., 2016; Geirhos et al., 2018).”
Benchmarking Robustness in Object Detection: Autonomous Driving when Winter is Coming (2019, background): “This drop in performance has previously been observed in image recognition models [e.g. Geirhos et al., 2018, Hendrycks and Dietterich, 2019].”
Using Videos to Evaluate Image Model Robustness (2019, background): “In contrast, human visual systems are robust to a wide variety of distortions, including color and noise distortions (Geirhos et al. (2018)).”
Using Pre-Training Can Improve Model Robustness and Uncertainty (2019, background): “Deep networks are notably unstable and less robust than the human visual system (Geirhos et al., 2018; Hendrycks & Dietterich, 2019).”
A Fourier Perspective on Model Robustness in Computer Vision (2019, methodology): “On ImageNet-C, we compare the robustness of the naturally trained model, AutoAugment (Auto) and the Stylized-ImageNet augmentation strategy (SIN+IN) (Geirhos et al., 2018a).”
Improving Robustness Without Sacrificing Accuracy with Patch Gaussian Augmentation (2019, background): “Research in neural network robustness has tried to quantify the problem by establishing benchmarks that directly measure it [24, 21] and comparing the performance of humans and neural networks [16].”
WILDS: A Benchmark of in-the-Wild Distribution Shifts (2020, neutral): “Examples include rotated and translated versions of MNIST and CIFAR (Worrall et al., 2017; Gulrajani and Lopez-Paz, 2020); surface variations such as texture, color, and corruptions like blur in Colored MNIST (Gulrajani and Lopez-Paz, 2020), Stylized ImageNet (Geirhos et al., 2018), ImageNet-C (Hend”
Biologically inspired sleep algorithm for artificial neural networks (2019, background): “Many studies in machine learning reported a failure of neural networks to generalize beyond their explicit training set [5].”
Location Attention for Extrapolation to Longer Sequences (2019, background): “However, many researchers have questioned the forms of generalization that neural networks exhibit, which signiﬁcantly diverges from human-like generalization (Lake and Baroni, 2017; Geirhos et al., 2018).”
Beyond neural scaling laws: beating power law scaling via data pruning (2022, methodology): “To this end, we use the model-vshuman toolbox [42] based on data and analyses from [43, 44, 45, 46, 47].”
Test-Time Training for Out-of-Distribution Generalization (2019, background): “Geirhos et al. (2018) show that training on images corrupted by Gaussian noise makes deep learning models recover super-human performance on this particular noise type, but cannot improve performance on another those corrupted by another noise type e.g. salt-and-pepper noise.”
The Notorious Difficulty of Comparing Human and Machine Perception (2020, background): “In other studies, the degradation of object classiﬁcation accuracy is measured with respect to image degradations (Geirhos, Temme, et al., 2018) or with respect to the type of features that play an important role for human or machine decision-making (Brendel & Bethge, 2019; Geirhos, Rubisch, et…”