We simulated different populations of visual neurons with on-center/off-surround spatial filters similar to bipolar cells in the retina [euler review]. The simulated population of N shifted but identical spatial receptive fields, each characteristic for one neuron, is shown in  Fig1a. We analyze the responses of these cells to \(48 \times 48\) Gaussian white noise input images. The weights of the receptive fields are normalized to produce neurally plausible mean response rates of 0.1 on these input images with added Poisson-like noise (zero centered normal with variance equal to the value of the absolute response).

In this simple example we reduce the GLM down to a linear regression problem solved analytically with ordinary least squares (OLS) and compare different forms of regularization, i.e. Lasso (L-1 norm) and Ridge (L-2 norm) both in their vanilla scikit-learn implementations (http://scikit-learn.org).