Our findings may be relevant to existing research since the computational and anatomical substrates of the Scintillating grid and other grid illusions are not fully understood (
Spillmann & Levine, 1971;
Wolfe, 1984;
Spillmann, 1994;
Schrauf et al., 1997). The Scintillating grid can be regarded as a stronger variant of the classic Hermann grid illusion, in which a somewhat similar effect is evident at the intersections of the bars in absence of any disks (
Hermann, 1870;
Spillmann, 1994;
Schrauf et al., 1997). The classical explanation of the Hermann grid by Baumgartner posits that the illusion is mediated by neurons in the retina with center-surround receptive fields (
Baumgartner, 1960,
1990). Baumgartner’s theory can, at least partially, apply to the Scintillating grid (
Thomson & Macpherson, 2018). However, more recent work using variants of the Hermann grid and the Scintillating grid stimuli suggest that Baumgartner theory alone is not sufficient to explain all grid illusions (
Spillmann & Levine, 1971;
Wolfe, 1984;
Spillmann, 1994;
Schrauf et al., 1997). For example, when the bars are distorted, the illusory perception is largely diminished for the Hermann grid and the Scintillating grid (
Schiller & Carvey, 2005;
Geier et al., 2008;
Levine & McAnany, 2008). Such results suggest the involvement of visual processing stages that are downstream to the retina, such as V1. To our knowledge, there is no direct electrophysiological evidence of this claim, which, indeed, may be nontrivial to obtain. Unlike the brain, the DNN models considered here permit easy access to the entire computation hierarchy. Analyzing the deviation magnitude along the computational hierarchy showed the largest deviations from monotonicity in the deep stages of computation, in both VGG-19 and ResNet-101 (
Figure 4). Although the computation in human and DNN is not immediately comparable, studies have suggested similarities between early DNN computation and human opponent-color and frequency-selective representations (in retina and in V1), while deeper DNN stages seem to be similar to deeper brain areas, such as V4 or IT (
Yamins et al., 2014;
Güçlü & van Gerven, 2015;
Kriegeskorte, 2015;
Cichy et al., 2016;
Eickenberg et al., 2017). Consequently, the results using the DNN metrics offered here are more consistent with a cortical (post-V1) origin of the Scintillating grid than an origin earlier in human visual processing, with the later stages of the DNN corresponding most closely to high-level cortical areas. We observed relatively higher layer deviation magnitudes for the sinusoid bars and no bars images in VGG-19 than ResNet-101 (
Figure 4), which both have some level of illusion under human observation (
Supplementary Figure S7). This difference between the two models may be reflective of VGG-19 generally being regarded as a better approximation for brain cortical processing than ResNet-101 (
Schrimpf et al., 2018).