Emission model

[ineslaranjeira]

• [x] Gaussian mixture models
• [x] test hyperparameter setting (I think I did this before and the effect on the fits was negligeable... - put this into a document?)
• [x] AR model
• [ ] SLDS
• [ ] Consider directly using code from Keypoint MoSeq: https://github.com/dattalab/jax-moseq/tree/5bdf11b92e3ddfd638d081233a9e3a36714c2227/examples
• [ ] Consider time-warped AR-HMM: https://openreview.net/forum?id=6Kj1wCgiUp_

[ineslaranjeira]

AR-HMM

• [x] Learn how to fit higher-order AR-HMMs
• [x] Add lag to the grid search together with number of states
• [ ] Understand how emissions model is fit?
• [ ] How to visualize dynamics/ How to visualize learnt parameters? (dynamics show where the system goes with an arrow, because it's linear - vector fields)
• [x] Plot inferred states overlaid with trial events like before
• [ ] How is prior set?
• [ ] Look at initialization - Kmeans;
• [ ] check initialization for Poisson HMM
• [ ] How can different lags be fit to different features (number of parameters will increase!)
• [ ] formally test how to capture two different wheel movement directions with 3-state HMM - model comparison