Figure 9.

Shuffling spike trains across trials distinguishes the network from a rate model. We explore how the decoding error varies as we decode the spiking output of the network where we replace an increasing number of individual neuron spike trains with those from different trial simulations. The parameters are the same as those used in Figure 4. a, Plots of the relative error (Eq. 47) as a function of the number of replaced spike trains for the network with the box function input. As the number of replaced spike trains is increased, so does the error. b, Plots an example network estimate (red) against the actual signal (blue) when no spike trains have been replaced. c, Plots the network estimate (red) against the actual signal (blue) when all 400 spike trains are taken from separate trials. Notice how replacing spike trains increases the variability of the estimate around its mean. For comparison, the relative error in b is 0.05, whereas in c it is 0.09. d–f, The same as a–c except that the OU stimulus is used. The relative error in e is 0.09, whereas in f it is 0.18.