Figure 1:

ON and OFF RF mosaics and their temporal kernels are predicted by efficient coding of natural videos. (A) Frames of natural videos $x\left(t\right)$ plus input noise $n_{\text{in}}$ are linearly filtered with the spatial kernels $w_j$ and then passed through one-dimensional temporal convolutions ${\varphi }_j$ followed by a nonlinearity, resulting firing rates $r_j\left(t\right)$ for each of $J$ RGCs. (B) The same calculations shown along the time axis, visualizing the temporal convolutions. (C) Examples of initial and optimized spatial filters, temporal filters, and nonlinearities: (left) fast OFF kernel, (right) slow ON kernel. (D) Unconstrained spatial filters $\left(J=160\right)$ learned center-surround shapes, about half of which are ON RFs. (E) Temporal filters $\left(J=160\right)$ using the parameterization (6) converged to four distinct clusters.