The Relevant ordinary differential equation identifier (Rodent) is a special kind of auto-encoder that enables the indentification of manifolds of generating ODE models.
The figure below shows a schematic of the Rodent:
The animations below show true data (dashed lines) and reconstructions (colored lines) and the distributions of each latent variable for a batch of harmonic signals. The Rodent needs a few more training iterations, but learns the simplest model for harmonic signals with only four relevant parameters.
The Auto-Encoder converges faster but its latent space is very complex:
