Hacker News new | past | comments | ask | show | jobs | submit login

Their derivatives, f^n(x) for arbitrary n, do not converge fine. Indeed, even n=1 is usually terrible in practice.

Fit a pleasant day-trip velocity curve with a NN and the acceleration would kill you.




Wouldn't that be mostly a consequence of the loss function used to compute the distance between the curve and the model output?




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: