Figure 1.

Model architecture for responses of a cortical neuron to visual stimuli, such as natural images. Light (ON pathway) and dark (OFF pathway) image regions are encoded as rectified responses of convolution with positive (light) and negative (dark) spatial gaussians, respectively. The output of this convolutional layer is then multiplied with a spatiotemporal filter of the same pathway that represents excitatory (red) and inhibitory (blue) weights for each spatiotemporal location. These linear-weighted sums are separately taken for each pathway and summed, followed by a half-power pointwise nonlinearity. A machine learning algorithm estimates the sizes of the parameterized Gaussian operators, and the two sets of spatiotemporal filters, for each of a series of latencies.