Using variational inference to perform weight uncertainty in the ODEFunc

rtqichen / torchdiffeq

Differentiable ODE solvers with full GPU support and O(1)-memory backpropagation.

MIT License

5.5k stars 909 forks source link

Using variational inference to perform weight uncertainty in the ODEFunc #116

Closed andreichalapco closed 3 years ago

andreichalapco commented 4 years ago

Hi,

I was wondering is there a practical way to perform variational inference to find the posterior distributions of the weights of our ODEFunc(neural network that defines the dynamics in a NODE). I'm trying an implementation with pyro, but I didn't got any good results and I also didn't find any research in this area. If you have some ideas, i will be very gratefull :). Thanks P.S: I know it is an Issue page, but I didn't know where to post the question.

rtqichen commented 3 years ago

There are quite a few papers on this now. You can either perform inference on the fixed-across-time parameters of the ODE (which can be done by pyro), or add an SDE prior and perform inference on entire trajectories (see torchsde).

Closing since it's not related to this library. Feel free to reach out via email if you want to discuss this.