This package allows fitting a latent circuit into the trajectories of the original continuous-time recurrent neural network (See Langdon et. al for details). The latent circuit fitting can be viewed as a model order reduction technique.
The main idea is to generate the trajectories x(t) with a small circuit (n units), embed it with matrix Q into the space of the original fitted trajectories y(t) (N units), and minimizie the differences Distance(Qx(t), y(t)), while simultaneuously ensuring that the small circuit performs the intended behavior well:
Here, the optimization is simultaneuously performed on both the embedding matrix Q and the recurrent connectivity of a small circuit w_{rec}.
One can start from the following [example Jupyter notebook](/jupyter/inferring latent circuit from CDDM RNN.ipynb)
This package relies on a [trainRNNbrain] package, make sure to download it and install it locally.