———— retweetledi
By far the clearest of the videos and by far the most jarring.
Eoin Higgins@EoinHiggins_
Doesn’t get much clearer than this
English
————
278 posts
@owl_posting “…Although we all prefer to use the good passport, sooner or later each of us is obliged, at least for a spell, to identify ourselves as citizens of that other place.”
English
@owl_posting I loved it. Reminded me of this quote: “ Illness is the night side of life, a more onerous citizenship. Everyone who is born holds dual citizenship, in the kingdom of the well and in the kingdom of the sick….”
English
Ask not why would you work in biology, but rather: why wouldn't you?
owlposting.com/p/ask-not-why-…
openai/gemini gave this essay an A- for evocative imagery. claude gave it a C- for being emotionally manipulative. both are probably right. i feel a little sick re-reading it
English
@Matthew_N_B I’d be curious too. But also, the way I see it, the “edge” here, though real, can be made vanishingly small, right? If the numbers selected are always astronomically large, then they always have a vanishingly small probability of being selected from the normal & P_win -> 0.5.
English
@owl_poster @norvid_studies I'm joking, the tone is lost on Twitter. Thank you very much for the kind words!
English
@dannymcneela For sure. I view it more as a learning platform, than a platform I can really develop a project on. Gotta keep the scope of the project minimal enough to be installable each time you run the notebook.
One plus side to not maintaining state is that it enforces reproducibility
English
@Matthew_N_B I keep trying to get myself to use Colab Pro, but I can't get past how annoying the interface is for the filesystem and terminal. The lack of persistent state and having to reinstall packages every time you reopen a notebook is a major downside for me.
English
@vallens I just saw this paper accepted at ICLR. Looks like an exploration of a similar idea: openreview.net/forum?id=yWiZa…
English
@langer_han I do understand that the kernel trick avoids ever having to deal with \phi(x) or any of this machinery in the reproducing kernel Hilbert space. In general, I'm trying to better understand how kernel methods work
English
@langer_han Thank you so much! It looks like I need to better understand \mathbb{E}[\phi(x) \phi(y) ] \otimes \mathbb{E}[\phi(x)] \mathbb{E}[\phi(y)] and how this generalizes a covariance matrix
English
@ed_kroc @MatthewBJane Take an arbitrary infinite sum of reals, x_1, x_2, ..., then \sum_{i=1}^\infty P({x_i}) < 1. I guess I'm still not sure what axiom is broken by the fact that there is NO infinite sum of reals that hits every real. Thus no sum for which \sum_{i=1}^\infty P({x_i}) = 1
English
@ed_kroc @MatthewBJane This makes intuitive sense, thank you! Which axiom exactly is broken? If we assign non-zero prob to every real number, then you can still take countable sums (you just can't sum over every real number), so you would be summing a subset.
English
@ed_kroc @MatthewBJane Thanks for the feedback! Can you clarify which sentences/statements are not quite right? Feel free to DM me. I'd like to make it as accurate/clear as possible!
English
@MatthewBJane @Matthew_N_B Those are VERY nice posts. My only significant quibble is with the intro to cont. RVs. Uncountability of the reals doesn’t mean we have trouble assigning probabilities to them (even all zero), it means we can’t add all them together one-by-one and expect things to work.
English
@Andrea1Mariani On the "social technology" side, I think we need to establish clearer language on what it is we are looking at and manipulating: mbernste.github.io/posts/cell_typ…
English
@Andrea1Mariani On the physical tech side we need deeper, less noisy assays (lots of progress being made there all the time). On the data analysis tech side, I think modeling the biophysical processes within cells is ultimately what is needed. Like this paper: cell.com/cell-systems/a…
English
@GorinGennady For better or worse, I don't know if DR methods developers see it that way. The orig t-SNE paper has a whole section devoted to "The crowding problem", which mostly addresses the visual clarity of the data points. Doesn't seem based in statistical modeling
English
@GorinGennady Under that conceptualization of what DR is, then I would completely agree these methods fall extremely short
English
@GorinGennady Very fair point. I would categorize that as an argument against using these methods in the specific field of single-cell analysis
English
@Matthew_N_B I have reservations about "seeing" the data when we know that the data are not the biological system (only a small fraction of the molecules are captured, and the DR methods do not explicitly represent that), and the conclusions are not drawn from individual cells' data.
English
@GorinGennady Oh gotcha. Yeah, agreed. I really like your line of work on modeling the biophysics of the cells and the process that generates the data. I need to sit down and work through the math, but that theory is ultimately what is needed!
English
@Matthew_N_B A sphere lives in continuous space; molecule counts live in discrete space. There are ways to connect these two concepts but nobody has done so to my knowledge (except in the context of generalized PCA).
twitter.com/GorinGennady/s…
Gennady Gorin@GorinGennady
@nmancuso_ @jcjray @JunhyongKim @lpachter That's more or less my intuition; I think that's a legitimate conceptualization (with the caveat that I really would want it to include the degenerate case of discrete cell types). But it would be nice to see people apply this idea, fitting instead of transforming data.
English
@GorinGennady ... policy of not allowing these plots in formal publications if they do mislead more often than inform.
English
@GorinGennady The goal is to "see" the data... but of course you can't really see it, so the goal is to see some aspect of the data that is preserved by the DR method however well defined that is. For supporting a conclusion, it can never be the only method. And I would understand the...
English