Imogen Binnie

The innocence of Astraea.

...by a team that included the first author 29 years earlier!

Because compensation attempts to remove the background autofluorescence, if a cell that's negative for the marker the immunofluorescence is binding to and also exhibits less autofluorescence than average, it will have a negative value after compensation (and this population will have a low mean both before and after compensation).

Overview from Handbook of Experimental Immunology

answers why distribution of fluorescence in flow cytometry data tends to be lognormal

Possible mechanisms of how serotonin increases the aversion to harming others.

Wow - I have not read many papers which are so... Biblical.

See Roederer, 2001, Spectral Compensation for Flow Cytometry: Visualization Artifacts, Limitations, and Caveats

https://onlinelibrary.wiley.com/doi/10.1002/1097-0320(20011101)45:3%3C194::AID-CYTO1163%3E3.0.CO;2-C

This is just a data format, essentially an array of detection events.

Summary of the 3 previous studies.

PD-1 blockade. Includes study of what differentiates successful from less successful treatment.

Also a PD-1 blockade

PD-1 blockade

Targeting 3 immune checkpoint molecules + a combination therapy yielded a meaningful survival difference.

Many tumors express B7-H1, which triggers apoptosis in activated T cells.

Shows that GBM relies on immune checkpoint molecules IDO, CTLA-4, and PD-L1. (IDO attracts regulatory T cells; CTLA-4 is expressed by T cells and CD80 on dendritic cells interacts with it during T cell activation; PD-L1 on a tumor binds to PD-1 on a T cell to tell the T cell not to kill the tumor cell.)

ie, a brain tumor cell line implanted in brain tissue.

GBM evades NK cell response in early tumor formation by overexpression of galectin-1.

GBM secretes periostin, a signaling protein involved in cell adhesion, wound healing, and the endothelial-mesenchymal transition. Overexpression of periostin also recruits immunosuppressive immune cells.

GBM expresses indoleamine 2,3 dioxygenase (IDO). IDO regulates the function & expansion of regulatory T cells. GBM recruits a ton of regulatory T cells that express GITR and seem to inhibit the immune response.

Many different tumors secrete the protein LDH5, which causes many healthy myeloid cells to produce NKG2D ligands, which causes NK cells to be less aggressive toward NKG2D ligands (down-regulates NKG2D receptors), which allows tumors expressing NKG2D ligands to get away scot-free.

"Interleukin 10 (IL-10) is a cytokine with a broad spectrum of immunosuppressive activity" and is produced by GBM cells.

transforming growth factor-\(\beta\)2 is produced by GBM and is immunosuppressing

CD45R/B220 appears prior to apoptosis (is the tumor shielding itself with a bunch of zombie T cells?)

this data is not right-censored

has the usual shortcoming that proportional hazard assumptions are biologically unlikely. we'll see if this becomes material later.

appears sensible when Delta & Omicron are the two VOCs. Alpha also has S gene dropout. BA.2 does not have this deletion, but that version of Omicron was not detected in the UK as of 12/7 (https://www.theguardian.com/world/2021/dec/07/scientists-find-stealth-version-of-omicron-not-identifiable-with-pcr-test-covid-variant)

U of Zurich

Columbia/Sloan Kettering. Computational/systems bio.

University of Zurich

This sidesteps the findings that IAT results don't seem to correlate with biased behavior. (https://qz.com/1144504/the-world-is-relying-on-a-flawed-psychological-test-to-fight-racism/ for a summary)

Fails to distinguish between a mathematical understanding of bias and the commonplace one.

This seems to represent a misunderstanding of what the embedding represents - news articles just need to quote more librarians who use 'she' and more philosophers who use 'he' to generate this. The writing need not be stereotyped - all that's necessary is for the world to be biased.

Was this really true - especially given that a key example embedding in the word2vec paper is about gender?

This seems like an ill-specified task? We'll see.

"Disturbing" implies some element of surprise, which seems unwarranted. (Doesn't make it less important, but the results aren't at all surprising based on the source texts.)

BART

ELMo

Small NLP lab (6-8)

From Berkeley NLP group.

Just got PhD in 2018.

4th yr PhD student

MIT, 2nd yr PhD student

This talks about the initialization for the transducers, but how is the domain model initialized? That is, where do the priors come from?

...because in its full generality measuring sentiment requires a complete understanding of social interaction, and is highly subculturally specific. This is not just an open research problem, it seems impossible without GAI.

BLEU score is an unintuitive metric here - wouldn't some way of measuring how well we discovered the codebook be better?

This is critical.

Ooh, this is interesting.

This seems like a lot of text on one of the simpler ideas in the paper?

From eq. 3 in the Kingma & Welling Auto-encoding Variational Bayes paper (https://arxiv.org/pdf/1312.6114.pdf)

Not generally true of an HMM - but maybe this is an assumption that is often made when doing inference? I don't know the usual HMM inference techniques.

But because the sequence is latent and corresponds arbitrarily to outputs, this doses not actually seem like a strong independence assumption to me.

This is nothing fancy, just marginalizing the distribution.

• The joint likelihood of the observed and latent sentences,
• given the parameters of the two transduction distributions
• ...is given by...
• the product over the m sentences we observe in the first domain of
• the probability of each observed sentence given the corresponding latent sentence and the transduction parameters
• times the probability under the language model that governs the second domain of the latent sentence
• all multiplied by
• the product over the n sentences we observe in the second domain of
• the probability of each observed sentence given the corresponding latent sentence and the transduction parameters
• times the probability under the language model that governs the first domain of the latent sentence.

General term for merged text of source and translation/transduction.

another plausible ancestor paper

The three dangerous statistical assumptions:

• the average is representative
• we have independence
• it's pretty much linear

Here Shannon establishes a relationship between the amount of noise and the maximum transmission efficiency under an error-correcting code that compensates for the noise with high confidence.

It is not totally clear what is in this category, is it. I wonder if the metaphor of "style" is limiting our imagination for how these models are used?

Hidden Markov Model.

A Hidden Markov Model in its full generality assumes very little about the data it generates - I wonder what this means in this context?

ICLR talk video here: https://papertalk.org/papertalks/4014

It's not clear to me how \(p_{\theta}\) is conditioned on the modified hidden state.

Is there a typo here? Is it not \(y_{\lt t}\) instead of \(y_{t-1}\)?

This is just regular Transformer training.

This is what we're aiming to maximize: the sum of the log likelihoods of the observations under the model parameterized by \(\theta\).

(Is there a bit of a problem here? In situations where x translates to multiple good y's, the ground truth probability p(y|x) is lower. Are we overweighting situations with fewer right answers?)

Training over a dataset of N paired sequences.

There's some question about to what extent this is a first-order problem and to what extent this is just difficulty in generalization. See: Quantifying Exposure Bias for Open-Ended Language Generation, He at al. & Generalization in Generation: A Closer Look at Exposure Bias, Schmidt.

Here's the key math. \(P(X|a) = P(a|X) * P(X) / P(a)\). Since we already have the constraint that a probability adds to 1, we can consider P(a) to be just a normalizing factor and we don't actually have to know/derive it.

Wide-ranging NLP contributions, very busy as a last author.

Other work includes: the breakthrough deep learning antibiotic discovery paper, several other molecular ML papers, and an algorithmic improvement for beam search (applied to machine translation). In 2nd or 3rd year of PhD as of 2021. Undergrad & MSc at MIT. Regina Barzilay = MIT adviser, Dan Klein = UCB adviser.

Alice's aces.

No - Bob wins if the river is a heart, and loses otherwise.

Alice does not have a draw to a straight.

Key takeaway: community members can identify people with high centrality in their social network.

This excerpt is similarly misleading, with that one "[while]" replacing a paragraph of text. (Again, yes Darwin was sexist, but these quotes misrepresent the text, and Darwin's stances).

I'm not going to defend Charles Darwin and I think we can all agree racism and sexism inhere to his worldview, but the quote is from The Descent of Man, not Origin of Species, and those ellipses are very misleading - they cut out more than 20 pages. As excerpted, it looks like Darwin is advocating genocide, but I don't think anyone would conclude that from the quotes in context.

Would be interesting to analyze the source of this. My first guess would be that it's generational, reflecting different usage patterns, rather than connectedness being "causal" (though I'm not positive I can nail down what the difference is).

Interesting - the linear model seems like a more naive approach to me. The approach described here doesn't become sensible until you tie in degree assortivity, which is not that intuitive, and not nearly strong enough to justify \(k^2\).

Who is this second-largest component!?

And has it been joined to the main component since 2011?

To annotate.

to read

to read

General observation about gradient descent unrelated to this paper - if you look at the partial derivative of any particular parameter, it's got two components, which, slightly metaphorically, correspond to how strongly it was activated and how much influence it had over the error. This differs from my introspective feeling about how my own learning works, where the understanding I'm most sure of, the part of my model which is most strongly activated, is not the thing that changes the most. It introspectively (and so unreliably) feels like I'm more likely to try to spin up a patch for the error - take an under-activated part of the model, shift its outbound connections to try to better match the shape of the error, and crank up the weight going into it. Gradient descent sort of only punishes mistakes rather than rewarding growth (though maybe that sense is just a consequence of the arbitrary choice of sign: minimizing loss vs. maximizing its opposite).

Basically, what if we looked for a highly activated cell that didn't affect the outcome much one way or another and see if we could make it affect the loss more and better? Once a cell gets a tiny weight going into it and tiny weights in its output stream, is there any hope for it to matter? How often does this sort of "self-pruning" zero-ing happen? Are these metaphors at all sensible?

But what else did you try??? More data on what doesn't work please!

typo -> "point of view"?

It seems like we could choose an arbitrary layer in the critic and use its input as this latent space - what choices of layer make this "perceptual"?

General GAN question - what happens if we make a different assumption about the z distribution?

The weights and biases in the first layer can scale and translate this along any dimension into any normal distribution, but what breaks if it's uniform instead? If we make our z distribution multimodal, does that help to learn multimodal domains better? Like ImageNet instead of faces, or non-centered objects, or scenes with more than one thing, all of which StyleGAN has a relatively hard time with?

Recent survey: https://arxiv.org/pdf/1902.06068.pdf

Should this be G(y,z; theta)?

Doesn't seem to be currently public as of 5/12/20.

Predicting the highest likelihood higher-resolution image vs. finding a probability distribution of higher-resolution images.

GAN originator, adversarial attacks.

Head of Computational Imaging team at Google. Looks like he was PI of a lab at UCSC that produced some very highly cited superresolution work in the 2000s.

Survey paper on image filtering.

The book on superresolution

Recent work focused on superresolution & denoising.

Google Scholar

Home page

Twitter

At Google since 2013. But on a research tear since ~2016-7.

Important work: Boundary Equilibrium GAN (https://arxiv.org/abs/1703.10717), MixMatch (semi-supervised learning algorithm) (https://arxiv.org/abs/1905.02249).

Google Scholar

Twitter

When I look at myself, when do I apply the schemer paradigm and when the kitten?

Ouch, is this true?

These seems like a subtle point, and it's not obvious to me at first blush that it should be so. Why couldn't this be irrational, say? There are an infinite number of families in this class of families.

Translation: The family is \(k|l-separated\) if any \(k-l\) elements of the ground set, and any \(l\) other ones, you can find a set in the family that has all of the \(k-l\) elements and none of the \(l\) other ones.

Seems like the notation would be better with this interpretation in mind. n|n-separated isn't as obviously silly as 0|n-separated. And then 1|1-separated would translate into the usual definition for separable, instead of 2|1.

If it's got every set of size \(l\), then no matter what \(k\) elements you pick, the set that contains exactly the first \(l\) is in there.

Is that the trivial example, or the only example? It's not the only example: {01234, 12345, 23450, 34501, 45012, 50123} is 5|4 separated (despite not having all the size 4 sets in it), but not 6|4 separated.

Because n|n separation doesn't say anything meaningful (there's nothing to check for absence).

Is it a Harvard convention that A is the treatment variable and Y is the outcome? Or a health care policy convention?

...in the course of founding epidemiology by isolating the cause of cholera to contaminated water supplies. (Compared cholera cases in houses supplied by two different water companies, before and after one switched the source of their water.)

See also this overview from the Columbia school of public health: https://www.mailman.columbia.edu/research/population-health-methods/difference-difference-estimation

Is this a Bayesian Network?

https://en.wikipedia.org/wiki/Bayesian_network

homework question: what makes diff-in-diff different from non-randomized trials (like, when trying to figure out the effect of surgical interventions) more generally?

This seems normal, why do we say so? Maybe the usual machinery accounts for people dropping out of the treatment group? Or are we preparing the way for later considerations about when there are a bunch of treatments of different populations that hit at different times?

units are the members of groups (control/treated).

is it? we could use a different estimator than OLS, like something that tries to compensate for heteroscedasticity (WLS, GLS), and still be looking for \(\beta\).

Oops, nope, units and groups are not the same thing. Units are in groups.

OK, this doesn't mean the simple thing I hoped it might.

I can make guesses, but it seems like it might be useful to say why we might choose one or the other, or under what circumstances incorporating the extra data is helpful or not. (It's not obvious to me that more observations necessarily helps - seems like it increases the odds that an exogenous factor is going to hit one group and not the other. And if some of these are ongoing factors and not shocks that dissipate, the performance of each group is decreasingly useful for estimating the other as time goes on.)

Yeah, this chart makes me sure that we're assuming that the treatment has no effect on the untreated group & the lack of treatment has no effect on the treated group.

From context, I'm guessing this paragraph is trying to teach me a difference between the notation used in this exposition & how it might 'usually' be used? I don't understand what new information was introduced.

This also seems like it ought to be part of the definition of \(Y^{a}\) rather than an assumption.

The distinction between parametric & non-parametric methods is pretty far over my head. Wikipedia gives me a very gross understanding, but how something could be in-between is DEFINITELY over my head.

Yeah! This seems much more straightforward! Why were we talking about a regression coefficient before? So confused.

this notation suggests that \(t=2\) is a pre-intervention time or \(T_{0}\), but below it stands in for the post-intervention period.

Know it's usual practice to do notation conventions first, but I'd prefer in an expository situation like this to see it defined as it comes up - I think I could pick up the definitions better in context. Would also help because there's more notation along the way than this table defines.

1) there must be a simpler way to write this.

2) couldn't this just be part of the definition of \(Y^{a}\) rather than an additional assumption?

I have some fundamentally different understanding about the nature of truth here that's making this hard to read for me - if we were to observe the potential outcomes, how would they still be potential? if we calculate the effect, how is it that we're still estimating the ATT and not calculating its actual value?

I know this is not a very complex formula, but it still seems unnecessarily complex for such a simple idea. It also obscures the fact that \(Y^{1}(2) | A=1\) is a fact that we have and \(Y^{0}(2) | A=1\) is an estimation. (right?)

magic number, ick.

These definitions I think I get, but they don't seem to fit the paragraphs above.

In the example I'd want the estimand to be "How much did California's inpatient spending change due to the new laws?"

I'm not sure why this is italicized or included in the sentence- when else would one expect treatment to have an effect?

This is sensitive to the time interval around \(T_{0}\) that we choose to include in the analysis, which makes it seem to me like it's not helping to answer the question of how much the intervention helped.

I'm definitely all turned around by now - we estimate a question that's answerable from the data by building a model that approximates the data?

just "function" would do, right?

animation could be clearer here, the nubs on the arrows are a little ugly and if the treated/control labels & graphs faded out, the brackets wouldn't have to fly so far.

I think the meaning here is ambiguous? We don't need a covariate to be continuous like it is in ANOVA, right?

This italics looks like estimand is being defined, but it isn't actually defined here, we just get an example. From context and Latin roots, I think it's "the thing we're estimating".

I'm also not clear how this question is statistical, or more statistical than the previous question - the change in the difference in spending in California vs. Nevada before and after the law changed seems like an observed fact.

Followup homework: diff-in-diff looks closely related to regression discontinuity - can it be viewed as a generalization of it? Are the observations in this explanation transferable?

hopefully this'll be clear from context later, but what is the output of \(Y\) exactly? A function from populations (eg CA, NV) to outcomes?

I get from context that \(T_{0}\) is when the intervention happens, but what's \(T\)?

this notation seems weird - why not \(T_{0}-n\)?

subtract? divide? depends on the context?

I'll probably find out in a bit.

How to cut a bagel into two equal halves that are linked together.

Probably not a useful conjecture for getting at FUNC.

"number of 1s"? Not sure that makes sense, any row or column can have at most 3 1s in it, all rows and columns are small if \(n\geq 5\).

One which can be non-zero only on the main diagonal, the diagonal above, and the diagonal below.

Pointer to Erdos-Ko-Rado and an extension by Hilton & Milner - could look at those to get ideas for tools?

Plus some history that's better covered in the Bruhn & Schaudt survey.

A couple of strengthenings of FUNC by looking at something even more general than a partial order, that seem not to be true (but maybe could be adjusted and re-opened in useful ways).

Don't think this direction is super likely to give leverage on FUNC because it moves things to a land with verrrry little structure.

I find this notation very confusing, but breaking it down:

• we've got a collection \(\mathcal{F}\) of n sets with a ground set 0-n
• each set contains its label
• no set contains any elements larger than its label

The rest of the question seems to imply the way to think about this is as a sort of truth table for an order? We have n items and each \(F_{i}\) tells you which elements are \(\leq i\).

It's reflexive because each \(F_{i}\) contains \(i\), antisymmetric because no \(F_{i}\) contains anything larger than \(i\), but it's not necessarily transitive. So it's not necessarily a partial order.

But you can take the partial order given by a lattice (or a intersection-closed family of sets) and represent it with one of these things.

this seems like we're losing a lot of generality! but "only" isn't necessary for the rest of the argument.

I don't follow this leap.

I don't see how this half of the inequality is motivated.

This isn't true under the assumptions we've made - the subcase 2 condition and the case 2 condition together means this isn't complete.

(Not that I'm clear why its completeness matters here. The previous statement, that everything in the powerset is present, is true, but because \(\Omega\) with each set intersected with some set is also union-closed, not because of completeness.)

the thrust of case I is correct - if we have a minimum set that overlaps with everything in \(\Omega - A_{k+1}\) and it also overlaps with \(A_{k+1}\), then the minimum set that overlaps with everything in \(\Omega\) is the same size.

minimum? only? \(g_{*}^{+}\) only has one value.

Does this mean to be an =?

True, but written strangely. (No need for the ceiling operator on the right side, or the log operator - it's just n+1.)

This doesn't follow. Each element of \(\Omega_{*}(B)\) is not, in general, distinct.

Think this inequality is the other way around by the earlier definition?

In other words, with m elements in the ground set, the most sets we can have in the family is 2^m. Yep!

Either there's an element or elements that are in every single set in \(\Omega\), or B is the whole universe of \(\Omega\).

In general, this is just B (so long as everything in the ground set is in some member of \(\Omega\), which it might as well be, or you could pick a smaller ground set).

We haven't defined what it means for this thing to be complete, though. It's a set, not an ordered tuple of sets.

I don't think this follows without a change to the definition of \(\lambda\).

This doesn't follow from the definition without some additional constraints. Some of the \(A_{i} \setminus C_{i}\) might coincide.

Ex: \((\emptyset),(1),(2),(12) \) reduces to \((\emptyset),(1)\) if you take \(C_{3}=C_{4}=(12)\).

\(\lambda\) answers: is union-closed family B formed by taking A and taking away elements from some of its sets?

If one union-closed family has one more set than another, then you can transform one into the other by adding or subtracting that set.

Union-closed families have a basis.

Other writers call this a basis set.

Or in other words, the labels that we stick on the ground set don't matter.

And this is "the set of all abundant elements in the union-closed family \(\Omega\)"

So this is just "all union-closed families".

The size of the smallest set in \(\mathcal{F}\).

I believe that v6 of this paper has unpatchable flaws.