Source-linked AI summary
Beyond accuracy: quantifying trial-by-trial behaviour of CNNs and humans by measuring error consistency
Robert Geirhos, Kristof Meding, Felix A. Wichmann
TL;DR
The paper addresses how to determine whether decision makers use similar strategies when accuracy alone cannot distinguish them. It introduces trial-by-trial error consistency and finds that CNNs agree strongly with one another but barely more than chance with humans, while CORnet-S behaves like a feedforward ResNet-50 in this analysis.
Problem
Accuracy alone cannot distinguish decision-making strategies, especially when humans and CNNs achieve near-ceiling visual recognition performance.
Method
The paper introduces error consistency, which compares whether two decision makers make errors on the same inputs rather than merely making the same number of errors.
Results
CNNs are remarkably consistent with one another, whereas human–CNN consistency is little above chance and CORnet-S behaves effectively like a standard feedforward ResNet-50.
Takeaways & Limitations
Error consistency suggests that human and machine vision still use very different strategies while providing a tool for quantifying progress in understanding complex systems.
Takeaways & Limitations
CORnet-S performs poorly on the reported behavioural error-consistency metric despite strong performance on neural predictivity and other benchmarks.
Abstract
from arXiv · showhide
A central problem in cognitive science and behavioural neuroscience as well as in machine learning and artificial intelligence research is to ascertain whether two or more decision makers (be they brains or algorithms) use the same strategy. Accuracy alone cannot distinguish between strategies: two systems may achieve similar accuracy with very different strategies. The need to differentiate beyond accuracy is particularly pressing if two systems are near ceiling performance, like Convolutional Neural Networks (CNNs) and humans on visual object recognition. Here we introduce trial-by-trial error consistency, a quantitative analysis for measuring whether two decision making systems systematically make errors on the same inputs. Making consistent errors on a trial-by-trial basis is a necessary condition for similar processing strategies between decision makers. Our analysis is applicable to compare algorithms with algorithms, humans with humans, and algorithms with humans. When applying error consistency to object recognition we obtain three main findings: (1.) Irrespective of architecture, CNNs are remarkably consistent with one another. (2.) The consistency between CNNs and human observers, however, is little above what can be expected by chance alone -- indicating that humans and CNNs are likely implementing very different strategies. (3.) CORnet-S, a recurrent model termed the "current best model of the primate ventral visual stream", fails to capture essential characteristics of human behavioural data and behaves essentially like a standard purely feedforward ResNet-50 in our analysis. Taken together, error consistency analysis suggests that the strategies used by human and machine vision are still very different -- but we envision our general-purpose error consistency analysis to serve as a fruitful tool for quantifying future progress.
1 Introduction1
The paper introduces trial-by-trial error consistency to compare decision-making strategies beyond accuracy. It motivates this analysis because aggregated accuracy and other metrics can obscure whether systems struggle with the same individual inputs.
- Contribution: Error consistency measures whether black-box systems make errors on the same inputs, providing a necessary but insufficient condition for similar strategies.The analysis applies across brains and algorithms, despite possible differences in implementation.
- Related work: The approach follows molecular psychophysics by targeting responses to individual stimuli rather than only aggregated responses or averages.This emphasis preserves information that metrics pooling across trials, stimuli, observers, or networks can lose.
- Motivation: Accuracy alone cannot distinguish strategies because systems may achieve similar performance while using different decision rules.The paper contrasts fine-grained trial-level error patterns with aggregated measures such as accuracy and confusion-matrix comparisons.
- Terminology: Similar outputs or errors are necessary but not sufficient for similar strategies, since different strategies can implement the same decision rule.Quicksort and Mergesort illustrate how distinct algorithms can produce identical outputs.
2 Methods
The method compares decisions on identical stimuli by separating observed agreement from agreement expected by chance and summarizing the difference with Cohen’s κ. It uses simulation-based uncertainty estimates and controlled stimuli to assess error consistency across perceptual systems.
- Observed vs. expected error overlap: Observed error overlap is the fraction of trials on which two decision makers give the same correct or incorrect response, compared with overlap expected from independent binomial observers.The expected overlap depends on the two observers’ accuracies and represents agreement attributable to chance.
- Uncertainty and bounds: Confidence intervals for overlap and κ are obtained by sampling independent-observer data, while analytical bounds provide additional consistency diagnostics.The sampling procedure addresses uncertainty in the chance-overlap estimate and avoids relying on the erroneous confidence interval discussed in the paper.
- Error consistency measured by Cohen’s kappa: Cohen’s κ discounts observed overlap by chance expected overlap to produce an accuracy-corrected error-consistency score.This yields a single behavioural measure that is disentangled from accuracy.
- Experimental methods: The analysis requires decision makers to evaluate the exact same stimuli in a regime that is neither perfect nor chance-level accuracy.The experiments compare humans and CNNs on shared psychophysical stimuli, including standard and out-of-distribution images.
- Experimental methods: The study evaluates ImageNet-trained CNNs from the PyTorch model zoo and the recurrent CORnet-S model against human responses.It also examines standard ImageNet and manipulated out-of-distribution images to probe model strategies.
3 Results
Error consistency reveals that CNNs agree strongly with one another but remain largely inconsistent with human observers, even as ImageNet accuracy improves. CORnet-S likewise shows no behavioural advantage over ResNet-50 on this trial-by-trial measure.
- 3.1 Comparing algorithms with humans: investigating whether better ImageNet models show higher error consistency with human behavioural data: CNN–human error consistency is close to chance for cue-conflict and line-drawing stimuli, with no improvement as ImageNet validation accuracy increases.Linear fits are nonsignificant for cue-conflict and line drawings: F(1,158)=0.086, p=.769, R^2=.001; and F(1,158)=0.478, p=.491, R^2=.003.
- 3.1 Comparing algorithms with humans: investigating whether better ImageNet models show higher error consistency with human behavioural data: CNNs show higher trial-by-trial consistency with one another than humans do, despite human observers achieving higher accuracy across experiments.The reported comparison includes silhouette accuracy of 0.75 for humans versus 0.54 for CNNs.
- 3.1 Comparing algorithms with humans: investigating whether better ImageNet models show higher error consistency with human behavioural data: Human–CNN algorithmic differences remain substantial across the investigated models, narrowing for silhouettes but persisting for cue-conflict, line drawings, and ImageNet images.AlexNet is reported as being just as error-consistent as recent models for the latter stimuli.
- 3.2 Comparing algorithms with algorithms: the “current best model of the primate ventral visual stream” behaves like a vanilla ResNet-50 according to error consistency analysis: CNN–CNN consistency is generally very high across architecture families, although models from the same family score higher on average.The architecture comparison used PyTorch models tested on edge stimuli, with 160 trials per observer.
- 3.2 Comparing algorithms with algorithms: the “current best model of the primate ventral visual stream” behaves like a vanilla ResNet-50 according to error consistency analysis: CORnet-S reaches κ=.066 with humans versus .068 for ResNet-50 and .331 for human–human consistency, while CORnet-S–ResNet-50 consistency is .711.AlexNet also exceeds CORnet-S at .080, and the figure comparison places CORnet-S only slightly above chance with most human observers.
4 Conclusion
Error consistency reveals that human and machine vision use substantially different strategies, while CNNs are highly consistent with one another and CORnet-S shows no behavioural advantage over ResNet-50.
- CNNs are remarkably consistent with one another irrespective of architecture.
- Human–CNN consistency is little above chance, indicating likely differences in their perceptual strategies.
- CORnet-S fails to capture essential characteristics of human behavioural data and behaves effectively like feedforward ResNet-50.
- Error consistency suggests that human and machine vision still use very different strategies.
- CORnet-S and ResNet-50 also score fairly similarly on a few Brain-Score metrics.
Broader Impact
The paper discusses both beneficial and harmful uses of error consistency, especially for interpreting algorithmic decisions and assessing claims about human-like AI.
- Error consistency can improve model interpretability, algorithmic accountability, and evaluation of claims about human-like AI.
- Quantifying differences between decision-making strategies may support more rigorous evaluation of algorithmic decisions.
- Measuring consistency between machines and humans could also be used to quantify progress toward machines that mimic human decision making.
- Quantifying differences between individuals can identify group-conform and outlier behaviour, with possible employment concerns if machines mimic human work.
Supplementary Material
The supplementary material provides terminology, statistical derivations, simulation procedures, model and Brain-Score details, and extended experimental analyses.
- Reproduction code and data for the results and figures are available online.
- The supplement derives bounds for c_obs and κ and simulates null-hypothesis confidence intervals for c_exp and κ.
- It documents Brain-Score and evaluated CNN methods and reports experiment-wide accuracies in Table 1.
- Extended results examine PyTorch models, CORnet-S versus ResNet-50, and relationships between Brain-Score metrics and error consistency.
- Figure SF.4 visualises qualitative error differences by showing stimuli particularly easy for humans or CNNs.
S.1 Terminology: “error consistency”
The term error consistency refers to testing whether decision makers agree on behavioural patterns beyond the agreement expected by chance.
- Error consistency assesses whether decision makers show significantly more agreement than expected by chance.
- The analysis considers consistency in both errors and correct responses.
S.2 Derivation of bounds for cobs and kappa given cexp
The derivation establishes theoretical bounds on observed consistency and κ for a given expected consistency, considering accuracy configurations above and below chance.
- The observed-consistency bounds are cobsmax = 1 −|pi −pj| and cobsmin = |pj + pi −1|.
- The analysis derives bounds for cobs and κ as functions of cexp, visualized in Figure 2.The bounds quantify the attainable consistency range beyond chance for specified expected consistency.
- When both accuracies are on the same side of 0.5, cexp lies in [0.5, 1] and the upper bound of cobs is always 1.The upper bound follows by choosing equal accuracies, pi = pj.
- When accuracies fall on opposite sides of 0.5, cexp lies in [0, 0.5[ and the case is point-symmetric to the right part.
S.3 Calculating 95% percentiles of observed overlap and kappa for the null hypothesis of independent observers given an expected consistency
The null distribution of observed overlap and κ is estimated by simulating independent observers across a dense accuracy grid and computing 95% percentiles.
- 95% percentiles for cobs and κ are obtained by simulating independent observers over a 4200 x 4200 grid of accuracy pairs.The simulated values are binned in 1% steps and quantified with R’s type-7 quantile function.
- The accuracy grid concentrates 66% of points in the upper and lower 15% because κ becomes unstable at high expected consistency.
- With 160 or 1280 trials, cobs has only 161 or 1281 possible values, limiting percentile accuracy near quantiles.The authors state that this limitation does not influence their findings.
S.4 Disentangling of Error consistency and Accuracy
The analysis separates κ from accuracy under independence: κ and cexp are uncorrelated for independent observers, while dependent observers can show correlations.
- For independent observers, κ and cexp are uncorrelated (r=-0.00015, p > 0.05), whereas κ remains constrained by accuracy.
- The simulations use 100.000 randomly drawn blocks for 160 and 1280 trials, with analytical bounds and 95% percentiles shown in Figure SF.1.
- κ shows no correlation with accuracy for independent observers, but correlations can occur for dependent observers.
S.5 Method details for Brain-Score and CNNs
The study compares human responses with CNNs from the PyTorch model zoo, including multiple architecture families and CORnet-S.
- Human responses were compared with CNN classification decisions from the PyTorch model zoo, spanning AlexNet, VGG, SqueezeNet, DenseNet, Inception, ResNet, and CORnet-S.The models were evaluated using torchvision version 0.2.2 implementations.
- Small result variations are expected because model implementations can differ across frameworks.
S.6 Error consistency of shape-biased models
Shape-biased training predicts more human-like error consistency, but CORnet-S lacks a human-like shape bias and behaves like ResNet-50 on these analyses.
- Shape bias predicts human-CNN error consistency for cue conflict stimuli: shape-based networks make more human-like errors.
- CORnet-S resembles ResNet-50 and lacks the human-like shape bias expected of an accurate model of human object recognition.
- CORnet-S behaves like ResNet-50 rather than human observers on parametrically distorted images.