Figure 2.

Recurrent motion model. (A) An Elman recurrent neural network with 750 input units all-to-all connected to the units of the hidden layer (d_ij). The hidden layer had 300 all-to-all recurrently and laterally connected hidden units (dashed lines, m_ik) that were all-to-all connected to the 26 output units of the RMM. Adjustable weights allowed the network to map motion inputs onto the speed tuned and direction selective response of the 26 recorded MT cells. (B) Example input-output. Low-pass filtered white noise patterns (range −1 to 1) shifted with 16°/s in the preferred motion direction over five time bins (bottom) is fitted to the measured response of the example MT cell to the preferred speed in the preferred motion direction (top, solid green line, the six other colors represent the non-preferred speeds).