Adapted gradient calculation missing?

[heinrichst]

Thanks for sharing your CTRNN implementation. However, it seems that the adaptation of the gradient calculation is missing (check Yamashita & Tani 2008, the proofs can be found in Hinoshita et al. 2011). Of course the default gradients would work to some extent as they would iteratively result in some local optimum. However, the weight matrices and thus the latent representation within the architecture would turn out to be quite different. I think the modified gradients are particularly crucial for the CTRNN in order to learn to abstract a functional hierarchy.

[Faur]

Thank you for pointing this out! It is definitely something I should check out. This project is not high-priority for me at the moment, so I don't when I will get around to it (but hopefully eventually).