Fig. 5: Incorporating photoreceptor adaptation enables CNNs to generalize across light levels.
Performance of (a) a conventional CNN model, and (b) the photoreceptor–CNN model. Each model was evaluated at three light levels (labeled below each box plot): high (column 1; yellow) and medium(column 2; orange), at which the models were trained, and low (column 3; red) which the models did not see during the training. The box plots indicate the median FEV across 37 RGCs, interquartile range (25th and 75th percentiles), and minima and maxima within 1.5 times the interquartile range. Outliers are plotted as individual points. Numbers at the top of each box plot are the median FEVs ± 95%c.i. c Performance of the conventional CNN model (blue color; same model as in a), and the photoreceptor–CNN model (green color; same model as in b) at all combinations of training and test light levels. For each column, the legend below the box plot panel shows the two light levels the models were trained at and the third light level at which it was tested (black outline). The box plots show the distribution of FEVs at this testing light level. Testing light levels were low (column 1), high (column 2), and medium (column 3). The photoreceptor–CNN model and the conventional CNN model showed statistically significant differences when tested at low light level (column 1; p = 5 × 10−7) and high light level (column 1; p = 1 × 10−7). p values were calculated by performing a paired two-sided Wilcoxon signed-rank test on the FEV distributions (N = 37) from the CNN and photoreceptor–CNN model at each testing light level. Source data are provided as a Source Data file.
