Skip to content

Milestones

List view

  • Improve documentation of source code We need to agree on a docstring format for auto documentation in read the docs or other

    No due date
  • Add more examples than GHMM in the source code and in the notebooks Success criterion: Have datasets implemented properly and corresponding Observation/Proposal/Transition models + illustration showing latent state

    No due date
  • Implement Auxiliary Particle filter etc Success criterion: run the Bootstrap PF and Auxiliary PF on a GHMM model with Systematic Resampling - compare likelihoods and (approximated) gradients

    No due date
  • Dependency on Sinkhorn for full completion Implement Reich regularisation plus gradient Success criterion: run the Bootstrap PF on a GHMM model - show table with likelihoods + gradients compared to Kalman and Systematic resampling

    No due date
  • Implementation of sliced Wasserstein point cloud optimisation + illustration notebook. Success criterion: run the Bootstrap PF on a GHMM model - show table with likelihoods + gradients compared to Kalman and Systematic resampling

    No due date
  • Implement neural nets F_theta(w, X) minimizing theta -> L(w, X, 1/N, F_theta(w, X)) For losses given by biased sinkhorn loss/unbiased sinkhorn loss/sliced wasserstein distance The neural net architecture should be permutation/rotation/scaling/translation equivariant (see e.g. PointNet). A good idea would be to generate w and X with a GAN approach: G_alpha -> w,X solving min_theta max_alpha L(G_alpha_0, G_alpha_1, 1/N, F_theta(G_alpha_0, G_alpha_1)) Success criteria -

    No due date
  • Provide notebooks for illustrative purposes on resampling techniques

    No due date
  • Improve unittests

    No due date
  • Implement all regularised resampling Success criterion: run the Bootstrap PF on a GHMM model - show table with likelihoods + gradients compared to Kalman and Systematic resampling

    No due date