Fig. 1: DCNN Modelling of orientation VPL.
a,b, A DCNN (a) is trained on an orientation discrimination task (clockwise, CW or counterclockwise, CCW) with Gabor stimuli embedded in different levels of image noise (b). c,d, Orientation discrimination accuracy is improved from pre-test (c) to post-test (d). e,f, Training induces a downshift of the threshold versus noise function (f), an effect that is qualitatively similar to existing human psychophysical results (e, corresponds to the 70.7% accuracy condition in fig. 1 of ref. 7). The absolute quantitative differences between e and f may be due to differences in the overall SNR or the number of layers and units between the human visual system and the DCNN. Data are presented as mean ± s.e.m, with error bars and error shadings in c and d representing the s.e.m. across four (n = 4) reference orientations.
