Dmitry Kobak

> Perhaps the negative KL values are the most puzzling of all. Aren't KL values during BH optimization computed using some expressions computed via BH? If so, then it's not...

> I'm just not sure what's the best way to detect that there are overlapping points in the initialization. Maybe something like ``` if initialization.shape[0] == np.unique(initialization.round(decimals=6), axis=0).shape[0] ``` can...

Somewhat fleshing out this workaround: ``` import numpy as np X = np.random.randn(1_000_000,2) X[1000] = X[2000] def jitter(X, precision=6): _, ind_unique = np.unique(X.round(decimals=precision), axis=0, return_index=True) ind_duplicates = np.setdiff1d(np.arange(X.shape[0]), ind_unique) if...

> I'm fine with using your fix as a temporary solution. Clearly, there is a bug somewhere, and it needs to be found. However, this isn't something that would happen...

Matlab version should be fine.

Hmm I am not quite sure what is going on in here, but in any case you should not include `[1 2]` in both marginalizations, so it should probably be...

3D embeddings are good for only one purpose: to create spinning and "fly-through" animations for public talks :-)

@renxwise Hmm. Can you give me any example of tSNE meaningfully/successfully used not for visualisation?

Yes, please do post a link here once it's ready. Sounds interesting.

Hey Toni, "pull request" does not mean closing the issue :-) I am reopening it, as it's clearly a bug.