SourceParsing neural dynamics with infinite recurrent switching linear dynamical systems
On the synthetic NASCAR task, where the true generative model is an RS LDS with k=4 states and recurrent dynamics, the SLDS model (which has no recurrence) attains a reasonable test log-likelihood (1331.35) but a very high dynamical mean-squared error (0.930) between its learned state-space flow and the true flow field, compared to 0.010 for the RS LDS. The SLDS sample trajectories do not resemble the true NASCAR-track dynamics. This was checked over 5 random initialization seeds.