Abstract
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.
Introduction
In the quest to build "human-like" artificial intelligence, researchers often use ImageNet accuracy as a proxy for progress. However, two systems can achieve identical accuracy using completely different strategies: one might struggle with lighting, while the other might be baffled by textures. If we want to know if a Convolutional Neural Network (CNN) actually processes visual information like a human brain, looking at the final score is insufficient. This paper addresses the "black box" problem of deep learning by analyzing the pattern of mistakes. If a model is a good proxy for the human visual system, it should find the same images "hard" or "easy" as humans do.
Measuring Error Consistency
The authors introduce error consistency, a metric derived from Cohen’s kappa ($\kappa$), to measure trial-by-trial agreement between observers. This approach distinguishes between observed overlap ($c_{obs}$), the frequency with which two observers give the same response, and expected overlap ($c_{exp}$), the agreement expected by chance given their individual accuracies. The resulting consistency metric, $\kappa$, is normalized such that 0 represents chance-level agreement and 1 represents perfect trial-by-trial consistency.
To ensure statistical rigour, the authors derived analytical bounds and used simulations to create 95% confidence intervals, allowing them to determine if observed consistency significantly exceeds what would be expected from independent systems.
Key Findings
The study evaluated 16 standard CNN architectures and human observers across four types of stimuli: standard ImageNet photos, cue-conflict images (shape vs. texture), edge drawings, and silhouettes.
The CNN Monoculture vs. The Human-AI Gap
Regardless of architecture, CNNs are remarkably consistent with one another, with $\kappa$ values ranging from 0.45 to 0.67. This suggests a shared "machine strategy" for vision that is distinct from biological processes. In contrast, while human-to-human consistency is relatively high ($\kappa \approx 0.32$), the consistency between CNNs and humans is significantly lower—often hovering just above zero. This indicates that humans and CNNs are likely using fundamentally different features to identify objects.
Accuracy vs. Human-likeness
Crucially, the research found that increasing a model's ImageNet accuracy does not necessarily make its behavior more human-like. As shown in panel (c) of the figure above, better models sometimes became less consistent with human error patterns. Furthermore, CORnet-S, a recurrent model often cited as the best model of the primate ventral visual stream, behaved almost exactly like a standard feedforward ResNet-50, failing to capture human behavioral characteristics.
Conclusion
Error consistency provides a necessary, though not sufficient, condition for claiming that two systems use the same strategy. While high consistency suggests a shared strategy, the low consistency observed here definitively proves that current CNNs and humans process visual information differently. The authors propose that error consistency should serve as a standard benchmark for the next generation of models, moving beyond simple accuracy toward behavioral metrics that ensure AI fails in the same predictable ways that humans do.
Summary updated