review contents of this paper about ncdssm HOT 7 CLOSED

Thanks for your early reply, I would be ready for Q&A in several days : )

from ncdssm.

Hi @666-will!

Thanks for your interest. Unfortunately, since the review process was confidential, I am not sure if I can reveal it publicly. That said, if you have any specific questions about the paper, I will be happy to answer them. You can also send me an email, if you prefer that.

from ncdssm.

So sorry for the delay of the issue.
First issue is about the ODE solver which might be used in the code:
In linear_step(mu: Tensor...tn,h,method: str = "rk4",), h is the gap between each time step.
Does it mean we just smoothly modeling irregular time series by setting a common divisor--- h ?

from ncdssm.

The second is about related work of NCDSSM.
In the pseudocode of GRU-ODE -Bayes, it simply apply GRU as a function to ODE-Solver, which called continuously modelling time series in that paper. More recently, Conti former that is published in nips2023 also has continuously model the irregular series.
What is the differences between GRU-ODE/conti former and NCDSSM?

from ncdssm.

In linear_step(mu: Tensor...tn,h,method: str = "rk4",), h is the gap between each time step.

h is the step size used for the ODE solver.

What is the differences between GRU-ODE/conti former and NCDSSM?

We have briefly discussed this in the related work section. GRU-ODE-Bayes uses a deterministic state which is updated via a Bayesian-inspired update step. In contrast, NCDSSM is an SSM with a stochastic state driven by an SDE and uses a principled Bayesian update to incorporate new observations. I still need to check Contiformer in detail but it looks like they're proposing a transformer-based model for continuous-time modeling and are not using a state-space formulation with a stochastic state like NCDSSM.

from ncdssm.

Thank you very much! Wish for your next amazing paper and have a good day : )

from ncdssm.

Thanks! :)

from ncdssm.