Source-linked AI summary
Generalisation in humans and deep neural networks
Robert Geirhos, Carlos R. Medina Temme, Jonas Rauber, Heiko H. Schütt, Matthias Bethge, Felix A. Wichmann
TL;DR
The paper asks how well humans and DNNs generalize object recognition across image degradations rather than only matching known distortions. Comparing human observers with standard and distortion-trained networks across twelve manipulations, it finds humans more robust overall, while DNNs excel on trained distortions but generalize poorly to unseen ones.
Problem
The paper examines how visual systems cope with new image degradations after learning to handle other distortions, extending robustness beyond fixed-distribution testing.
Method
The study compares human and DNN object-recognition robustness across twelve image distortions, including standard ImageNet networks and networks trained on distorted images.
Results
Humans were more robust than standard DNNs across nearly all tested manipulations, whereas distortion-trained DNNs performed extremely well on trained distortions but failed to generalize strongly to unseen distortions.
Takeaways & Limitations
Robust generalization across changing distortion distributions remains a central challenge for DNN vision systems and a target for lifelong machine learning.
Takeaways & Limitations
The study does not resolve how DNNs can overcome this generalization challenge under non-i.i.d. conditions, leaving this to future investigations.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper asks whether DNNs generalise across image degradations as robustly as humans, extending beyond conventional i.i.d. generalisation and transfer learning.
- Motivation: Human vision generalises across illumination, weather, and other input-distribution changes, including previously unfamiliar conditions.This robustness motivates comparisons with machine vision systems intended for settings where the test distribution is not known during training.
- Generalisation question: DNNs achieve high accuracy on i.i.d. test data and can transfer features across loosely related tasks, but transfer learning requires additional training.The paper focuses instead on coping with new image degradations after learning other distortions.
- Study aim: The experiments target object-recognition robustness when signal strength decreases under image distortions.The comparison uses classification performance and error-pattern differences as signal quality weakens.
- Study aim: The study compares ImageNet-trained GoogLeNet, VGG-19, and ResNet-152 with humans across twelve distortions, then trains networks directly on distorted images.The two experiment sets assess robustness to unseen distortions and whether distortion-based data augmentation transfers across conditions.
2 Methods
The methods establish a controlled forced-choice comparison between humans and DNNs using mapped 16-class ImageNet categories, twelve image manipulations, and varying signal strengths.
- Paradigm and task: The study uses a forced-choice image-categorisation paradigm to compare human observers and DNNs on matched entry-level categories.Fine-grained ImageNet outputs were mapped to 16 coarse classes by summing probabilities for corresponding fine-grained categories.
- Paradigm and task: Human presentation was limited to 200 ms followed by a 200 ms 1/f-spectrum noise mask to reduce feedback-related processing.This design addresses the difference between feedforward DNN inference and recurrent connections in human vision.
- Participants and models: Human data were collected from three to six observers per experiment, with observer counts varying across distortion experiments.Observers had normal or corrected-to-normal vision, and the DNN comparison used pretrained VGG-19, GoogLeNet, and ResNet-152 models.
- Stimuli: Twelve experiments applied image distortions across signal strengths ranging from undistorted full signal to distorted weaker signal.Manipulations included colour, greyscale, contrast, filtering, phase noise, power equalisation, opponent colour, rotation, Eidolon distortions, and additive noise.
- Distortion-trained networks: Distortion-trained networks used 16-class ImageNet, class-weighted losses, and on-the-fly augmentation with selected manipulation combinations.The training experiments compared networks trained on single distortions, uniform noise plus another distortion, or nearly all distortions.
3 Generalisation of humans and pre-trained DNNs towards distortions
Across twelve degradations, humans remained more robust than pretrained DNNs as signal weakened, while DNN responses became increasingly category-biased.
- Accuracy and error patterns: Human observers were more robust than pretrained GoogLeNet, VGG-19, and ResNet-152 for nearly all distortions as signal strength decreased.The human advantage was small for low contrast, power equalisation, and phase noise, and larger for noise, filtering, rotation, and Eidolon distortions.
- Accuracy and error patterns: DNN response distributions became increasingly biased toward certain categories at weaker signals, whereas human responses remained approximately balanced across 16 classes.Response-distribution entropy was used to quantify this bias.
- Interpretation: The results indicate that DNNs have greater difficulty than humans generalising to weak signals across varied image distortions.The human advantage occurred despite the tested manipulations not necessarily matching people’s exact prior visual experience.
- Follow-up experiment: The findings motivated a second experiment batch that trained networks directly on distorted images to examine whether training exposure explains the robustness gap.This tested both direct performance on distortions and transfer to other distortions.
4 Training DNNs directly on distorted images
Networks trained on individual distortions can outperform humans on the distortion seen during training, but this specialised success does not transfer reliably to unseen distortions. Broader distortion training offers limited generalisation benefits, while longer training may explain some gains.
- Specialised networks consistently surpassed human observers on the exact distortion used for training.This suggests standard architectures and training procedures can solve distortions under i.i.d. train/test conditions.
- Training on one distortion generally failed to improve performance on other distortions and sometimes reduced it relative to vanilla ResNet-50.Salt-and-pepper and uniform-noise performance remained near chance, even when networks were trained on the other noise type.
- Models trained with uniform noise plus another distortion improved on some trained and untrained distortions, but the gains may largely reflect 200 rather than 100 training epochs.The uniform-noise model also improved when trained longer without adding another distortion.
- A single model achieved high accuracy across eight trained distortions, yet reached only 11 to 14% on left-out uniform and salt-and-pepper noise.This remained much closer to chance, approximately 6%, than to specialised-network accuracy above 70%.
- The findings indicate that distortion-based data augmentation alone may be insufficient to overcome DNN generalisation failures.The authors point toward mechanisms such as neural normalisation and shape bias as possible future avenues.
5 Conclusion
The study finds marked differences between human and DNN processing of degraded object images and shows that training on distortions does not resolve unseen-distortion generalisation. Its behavioural dataset provides a benchmark for investigating more robust machine and brain models.
- 82,880 psychophysical trials compared human and DNN object-recognition robustness across twelve image distortions.The comparison used ResNet-152, GoogLeNet, and VGG-19 alongside human observers.
- DNN classification performance declined rapidly with decreasing signal-to-noise ratio, while human and DNN error patterns diverged progressively as signals weakened.
- Training on distorted images enabled DNNs to cope with the exact trained distortion but left strong generalisation failures on previously unseen distortions.
- Because possible distortions are effectively unlimited, training on all distortions is not feasible.
- The freely available behavioural data are presented as a benchmark for improving DNN robustness and motivating investigation of brain mechanisms underlying human robustness.
Supplementary material
The supplementary material documents the experimental procedure and controlled presentation conditions, including trial timing, response behaviour, stimulus sampling, and display specifications.
- The supplementary analyses include network uncertainty as a function of signal strength and accuracy across conditions for distortion-trained networks.
- Each trial used 300 ms fixation, 200 ms image presentation, a 200 ms 1/f noise mask, and a 1500 ms response interval.Observers clicked the category that best matched the image.
- Observers could change their category choice during the fixed 1500 ms response interval, with the final click counted as the answer.
- Twelve experiments randomly selected 70 to 80 images per category without replacement, counterbalancing conditions and randomising presentation order.Stimulus selection was individualised for each participant.
- Stimuli were presented centrally as 256×256-pixel images subtending 3×3 degrees on a 1920×1200, 120-Hz monitor.A chin rest helped maintain a constant viewing position.
Observers & pre-trained networks
The study used a small set of observers, established pretrained ImageNet networks, and a 16-category object task presented through controlled psychophysical trials.
- Observers: Observer samples ranged from three to six participants across the colour, contrast, noise, filtering, phase, power-equalisation, and eidolon experiments.Participant demographics and experiment-specific counts were reported separately.
- Pre-trained networks: The pretrained-network analyses used GoogLeNet, VGG-19, and ResNet-152 implementations and weights from TensorFlow-Slim.
- Trial procedure: The trial displayed a central fixation square, a 200 ms image, a 1/f noise mask, and a category response screen.
- Experiment structure: Table 1 reports experiment-specific trial counts, conditions, practice and main blocks, per-condition trials, and durations without breaks.
- Pre-trained networks: Images were fed through the networks using a single feedforward pass.
Categories and image database
The study combined ImageNet’s large image pool with MS COCO’s entry-level categories, then applied controlled preprocessing and multiple image distortions to create experimental stimuli.
- Categories and image database: 16 non-ambiguous entry-level categories were retained by mapping ImageNet labels to MS COCO categories through WordNet hypernym relationships.ImageNet supplied the images, while MS COCO provided categories such as dog rather than German shepherd.
- Categories and image database: ImageNet images were filtered by excluding greyscale images and images smaller than 256 × 256 pixels before center cropping.The preprocessing exclusions removed 1% of images for greyscale and 11% of non-greyscale images for insufficient size.
- Categories and image database: Contrast, uniform noise, colour, spatial filtering, Fourier-domain manipulations, rotations, and eidolon transformations generated the tested stimulus conditions.Uniform noise was added pixelwise after scaling images to 30% contrast, while other experiments varied filter, phase, rotation, or eidolon parameters.
- Categories and image database: Opponent-colour stimuli were designed to reverse perceived colours while preserving luminance using the DKL colour space and measured monitor characteristics.The procedure incorporated monitor luminance measurements and cone-activity calculations before transforming RGB values.
- Categories and image database: Eidolon stimuli were generated by varying reach and coherence while holding grain constant at 10.0.The toolbox function produced distorted versions of the original images from these three parameters.
Experimental modifications
The second experimental batch modified masking, image resizing, file format, software version, and related procedures, while the figures compared network uncertainty and cross-condition performance.
- Experimental modifications: The second batch enhanced the pink noise mask by multiplying pixel values by four and clipping values to the [0, 1] range.The change was intended to improve the mask’s masking effect.
- Experimental modifications: Human and DNN stimuli were downsampled to 224 × 224 in the second batch instead of giving humans 256 × 256 images and DNNs a 224 × 224 centre crop.This change ensured that humans and DNNs saw the same image dimensions.
- Experimental modifications: The study examined JPEG-versus-PNG storage because JPEG’s lossy compression may introduce image artefacts.ImageNet’s JPEG format motivated the initial choice of JPEG files.
- Experimental modifications: Python Version 3.5 replaced Version 2.7 for image preprocessing in the second batch.
- Experimental modifications: A six-month gap separated the batches, and human participants received no correctness feedback during the experiments.The time gap was intended to minimise memory effects.
- Experimental modifications: Figure 7 compares human observers with vanilla, specialised, and all-distortions networks across classification accuracy and response distribution entropy.The specialised networks were trained on one distortion, whereas the all-distortions network was trained on multiple distortions simultaneously.
Error bars & entropy
Accuracy ranges were reported across observers or matched DNN runs, while entropy was averaged across individual human participants to preserve individual differences.
- Error bars & entropy: Accuracy error bars show the range across different observers rather than the standard error of the means.For DNNs, seven data-subset runs matched the number of images per category and condition seen by one human observer, and their accuracy range was reported.
- Error bars & entropy: Human response distribution entropy was calculated as the average of individual participants’ entropies rather than from aggregated human trials.Aggregation could allow individual differences to cancel and would produce higher human response distribution entropy.
Prediction uncertainty
Network prediction uncertainty was assessed with entropy over the 1,000 ILSVRC classes, revealing that stronger distortions do not always produce less confident predictions.
- Prediction uncertainty: Uncertainty generally increases with distortion strength, but all networks become more certain again at some higher uniform-noise and eidolon levels.
- Prediction uncertainty: ResNet-152 becomes more certain under stronger low-pass distortions and is consistently more confident than GoogLeNet and VGG-19.
- Prediction uncertainty: Some networks restrict predictions to a few highly confident classes as the input signal worsens, indicating poorly represented uncertainty in these cases.The reported pattern is consistent with previous reports about uncertainty in standard discriminative DNNs.
Network training results across all distortion conditions
Across all stimulus levels, human performance remains better than vanilla ResNet-50, while distortion-trained networks achieve extremely high accuracies. Specialised-Net is generally comparable to or slightly better than All-Distortions-Net, with human-like response distributions even under difficult noise.
- Human performance exceeds vanilla ResNet-50 across the distortion experiments.
- Specialised-Net and All-Distortions-Net reach extremely high accuracies across stimulus levels.
- Specialised-Net is generally on par with or slightly better than All-Distortions-Net.
- Both distortion-trained networks show largely human-like response distribution entropy, near 4 bits, even when accuracy is low.This pattern is reported for difficult uniform and salt-and-pepper noise conditions.
- Figure 9 confusion matrices show correct categories by columns, classification decisions by rows, and failures to respond in the top row.
Aggregating probabilities from coarse classes
The paper replaces ad hoc probability summation with a principled method for changing class priors at test time and aggregating fine-grained classes into coarse classes. The method improves measured accuracy and entropy, although the paper’s conclusions remain unchanged under the earlier aggregation choice.
- Limitations and recommendations: The aggregation choice affects experimental accuracy and response-distribution entropy, but none of the paper’s conclusions depends on using the earlier ad hoc choice.The authors recommend the optimal aggregation method for future experiments.
- Changing priors: Discriminative DNNs can adapt their posterior predictions to an arbitrarily chosen test-time class prior without retraining.The new posterior uses the network’s old posterior together with the old training prior and the new prior.
- Posterior derivation: The new posterior is obtained by multiplying the old posterior by the new-to-old prior ratio and renormalizing.The normalization constant is fixed for a given input and need not be calculated explicitly.
- Coarse-class aggregation: For coarse classes, the method represents each coarse class as a disjoint set of fine-grained classes, which may vary in size.The paper uses sixteen coarse classes covering 207 fine-grained ImageNet classes, while ignoring the remaining 793 classes.
- Coarse-class aggregation: Coarse-class probabilities are computed, up to normalization, as averages of the corresponding fine-grained probabilities, and the largest average determines the decision.This follows from the paper’s optimal aggregation formula rather than the earlier ad hoc summation choice.
- Empirical effects: Using the new aggregation method increased clean-image accuracy from 98.3% to 98.8% and changed noisy-image accuracy and entropy by up to 10% and 1.8 bits.Depending on the condition, the method also produced overestimates of up to 4% and 1.2 bits.