Examples of learning dynamics and forward/inverse models

[tcstewar]

I've seen a number of talks of online learning of dynamics (i.e. observe a lorenz attractor, do online learning of weights, and result in a system that produces the lorenz dynamics). For example:

https://arxiv.org/abs/1902.02875

I know @arvoelke and @celiasmith have implemented this in nengo, but I don't think we've done a broad set of examples showing this. There is an example of FORCE learning here: https://github.com/arvoelke/nengolib/blob/master/docs/notebooks/examples/full_force_learning.ipynb

There's also this sort of work where dynamics learning is used to learn the forward dynamics of a physical system:

https://elifesciences.org/articles/28295

and the inverse model:

http://proceedings.mlr.press/v80/gilra18a.html

[arvoelke]

Agreed! This would also make an excellent opportunity for a review paper.

[tcstewar]

It's also worth noting that the elife paper used nengo to implement the model, and the code is available at https://github.com/adityagilra/FOLLOW