notes.txt

To-implement, in order of easiness:

1. get overexpressed genes for each cluster
    - DONE
2. more visualization tools - heat maps, gene expression in scatter plot 
2. uncertainty sampling by cluster assignment entropy
3. re-labeling/re-initialization of uncurl
    - including cluster merging/splitting
    - TODO
4. improved distribution selector, using p-values?
    - try out various things: implementing Chi-square test?
    - implement that crazy p-value thing from that one paper?
5. improved gene selection methods??? bring back robust uncurl???
    - use Poisson test/c-score test on cluster subsets to identify differentiating genes?
6. hierarchical uncurl
    - binary splitting w/uncurl, soft assignments based on entropy
7. picking K

Uncurl question: how many iterations do we actually need???

to investigate:
- gene selection while doing cell state estimation - use Poisson Test? (might fail for very sparse genes). Use c-score? (too ad-hoc).

- streaming uncurl? in terms of returning/yielding partial results during the optimization process, not in terms of operating over data streams (although that would be interesting too? streaming NMF?)

regularization in uncurl