Fig. 4: Incorporating photoreceptor adaptation enables CNN to predict responses of an example RGC at a light level different from those at which it was trained. | Nature Communications

Fig. 4: Incorporating photoreceptor adaptation enables CNN to predict responses of an example RGC at a light level different from those at which it was trained.

From: Biophysical neural adaptation mechanisms enable artificial neural networks to capture dynamic retinal computation

Fig. 4

a White noise movie at three different light levels: high (column 1; yellow), medium (column 2; orange) and low (column 3; red). b Recorded response (normalized spike rate) of an example RGC (gray lines) to white noise movie at the three different light levels in (a) (columns). Inset above the right column overlays a segment of the responses at the three light levels to directly compare response kinetics. c Responses predicted by a conventional CNN model (colored) at each light level in (a) (columns). FEV values above each trace quantify the performance of the model for this RGC at the corresponding light levels. d Same as in c but for the proposed photoreceptor–CNN model. Models were trained on data at high 30 R*receptor−1 s−1 (column 1) and medium 3 R*receptor−1 s−1 (column 2) and evaluated at low 0.3 R*receptor−1 s−1 (column 3) light level. Source data are provided as a Source Data file.

Back to article page