Supplementary Figure 11: The LFADS generator.
From: Inferring single-trial neural population dynamics using sequential auto-encoders
The generative LFADS model is a recurrent network with a feed-forward read-out. The generator takes a sampled initial condition, \({\hat{\mathbf g}}_0\) and a sampled inferred input, \({\hat{\mathbf u}}_t\), at each time step, and iterates forward. At each time step the temporal factors, ft, and the rates, rt are generated in a feed-forward manner from gt. Spikes are generated from a Poisson process, \({\hat{\mathbf x}}_t \sim {\rm{Poisson}}({\mathbf{x}}_t|{\mathbf{r}}_t)\). The initial condition and input at time step 1 are sampled from diagonal Gaussian distributions with zero mean and fixed chosen variance. Otherwise, the inputs are sampled from a Gaussian auto-regressive prior.
