Fig. 8: Alignment between human and CNN responses in out-of-distribution scenarios. | Nature Communications

Fig. 8: Alignment between human and CNN responses in out-of-distribution scenarios.

From: Improved modeling of human vision by incorporating robustness to blur in convolutional neural networks

Fig. 8

A Classification accuracy for standard (red), weak-blur (blue) and strong-blur CNNs (purple) based on aggregated performance for 17 out-of-distribution datasets provided by Geirhos et al. (2021). Note that 1 of 17 conditions involved blurry images, which was not out-of-distribution for the blur-trained CNNs. B Accuracy difference between humans and CNN models. C, D Consistency of responses and error responses between humans and CNNs; higher values indicate better human-AI alignment, with gray bars indicating human-to-human consistency. Source data are provided as a Source Data file.

Back to article page