IC-1087SLDS produces poor dynamical accuracy on the NASCAR task because it lacks recurrent switching

Victor Geadah, International Brain Laboratory, Jonathan W. Pillow

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.

Evidence
correlational
Key metric
NASCAR (k=4, 5 seeds): SLDS test LL 1331.35, dyn MSE 0.930; RS LDS test LL 1628.57, dyn MSE 0.010
Caveat
The NASCAR task is synthetic and was specifically designed to test the inclusion of recurrence; the gap may be larger on this task than on other data.
Model
SLDS
Concepts
Failure mode
Methods
Sequential Monte Carlo [eval]
Related work
Fox et al. 2010 (SLDS / Markov switching processes) [builds-on]
Related findings
IC-1086
Extraction
automatic-extraction