Approximate cost of training?

[bobinwobbin]

The paper was nicely detailed about most specifications but I can't really find any kind of informations like total training time or cost for that matter and it would nice to get an approximate feeling for it

[lukedarlow]

Hello, and thank you for your question.

You are right that we don't dive into very much detail. There are settings for training iterations in the appendices of the paper, just in case you missed those.

The training time is unfortunately longer than what you would expect from a purely feed-forward system. Naturally, this is expected owing to the pure sequentiality of the internal recurrence. We did our best to speed up as many parts of the process as possible (e.g., "Appendix K: Recursive computation of the synchronization matrix" in the paper).