Figure 4.

(A) BagNet-33's probability of correct class for decreasing crops: The sharp drop when the image becomes too small or the resolution too low is called the “recognition gap” (Ullman et al., 2016). It was computed by subtracting the model's predicted probability of the correct class for the sub-MIRC from the model's predicted probability of the correct class for the MIRC. As an example, the glasses stimulus was evaluated as $0.9999-0.0002=0.9997$. The crop size on the x-axis corresponds to the size of the original image in pixels. Steps of reduced resolution are not displayed such that the three sample stimuli can be displayed coherently. (B) Recognition gaps for machine algorithms (vertical bars) and humans (gray horizontal bar). A recognition gap is identifiable for the DNN BagNet-33 when testing machine-selected stimuli of the original images from Ullman et al. (2016) and a subset of the ImageNet validation images (Deng et al., 2009). Error bars denote standard deviation.