orange3xchicken

You probably couldn't find much evidence re Krugman + Charlatan, because he's not. He's an accomplished economist who's influential and respected among the academic community. He's won two of the most prestigious awards in economics (Nobel + John Bates Clark). He's had some peculiar arguments in the past, which were dumb imo, but nothing "Charlatan"... Science is hard and economics is unique. Krugman is known for doing post-mortems and admitting when his ideas are off. I also think that being recognized by the academic community and his peer economists is a reasonable signal that he's not an idiot.

More recently he's been a bit more into hackish punditry and is prob. most famous among regular people for his somewhat abrasive nyt column. I think your idea about and API is not ideal. Is it enough to trust some randos on hacker news to identify Charlatans?

At least on the quant side, I think the typical sentiment is that most researchers aren't interested in ops / developing infrastructure / curating datasets.

It sounds like you aren't really interested in a rational discussion by the second half of your post, but the typical arguments (incl in the post) for are that market makers reduce inefficiencies in the market & provide liquidity that significantly reduces the bar (i.e. make trading cheaper) for retail investors (like you or me) to trade.

I think it is generally accepted that society does benefit from a modern and efficiently run market. Whether or not automated market makers contribute to this could be up for debate, I guess.

I know, I've been waiting so long for this! I will not miss having to render and reference pngs of equations / api requests.

I think K is usually interpreted as the # of gaussians from which the data has been assumed to be sampled. Not immediately related to # of latent variables unless you invoke kernel kmeans or something like laplacian eigenmaps/diffusion maps.

The term "labeled graph" just means a graph with each node labeled differently (but arbitrarily). It just allows for reasoning & enumerating the vertex set. It's a typical assumption to make in the context of graph isomorphism.

It doesn't relate to machine learning (which is what I assume you mean).

That's really cool. These days similar strategies based on graph coarsening and vertex ordering are really popular for improving the sparsity pattern of preconditioners- e.g. incomplete lu decompositions for iterative linear solvers.

Just last night I noticed that the zoom program on my mac has been suffering from this bug, which is ~4 months old. Both are up to date.

https://community.zoom.com/t5/Meetings/Why-is-the-Zoom-app-l...

I'm less familiar with poisoning, but at least for test-time robustness, the current benchmark for image classifiers is AutoAttack [0,1]. It's an ensemble of adaptive & parameter-free gradient-based white-box and gradient-free black-box attacks. Submitted academic work is typically considered incomplete without an evaluation on AA (and sometimes deepfool [2]). It's good to see that both are included in ART.

[0] https://arxiv.org/abs/2003.01690

[1] https://github.com/fra31/auto-attack

[2] https://arxiv.org/abs/1511.04599

A new subfield of adversarial ML that considers similar challenges to adversarial NLP: topological attacks on graphs for attacking graph/node classifiers.

Both problems (NLP & graph robustness) are made much more challenging compared to adversarial robustness/attacks on image classifiers due to their combinatorial nature.

For graphs, canonical notions of robustness wrt classes of perturbations defined based on lp norms aren't so great (e.g. consider perturbing a barbell graph by removing a bridge edge- huge topological perturbation, but tiny lp perturbation!)

I think investigating robustness for graph classifiers should also help robustness for practical nlp systems and visa-versa. For example, is there any work that investigates robustness of nlp systems, but considers classes of perturbations defined on the space of ASTs?

Recommend Russel's Human Compatible. Three principles to guide AI development:

1. The machine's only objective is to maximize the realization of human preferences.

2. The machine is initially uncertain about what those preferences are but then is.

3. The ultimate source of information about human preferences is human behavior.

Maybe Bosch? Prof. Zico Kolter from CMU is a chief scientist associated with them, and his group does a lot of really good work in the ml verification space (e.g. the first randomized smoothing and the Wong & Kolter certificates).

For anyone interested in elegant implementations of state of the art algorithms for verification, there is a nice library in Jax:

https://deepmind.com/research/open-source/efficient-and-tigh...

Not sure if this is what you mean, but I found an ongoing Wikimedia research project & preprint:

https://meta.wikimedia.org/wiki/Research:Link_recommendation...

https://arxiv.org/abs/2105.15110

Not arguing with your point, but the blame shouldn't entirely be attributed to the research community. It's easy to read about the various parasitic practices conducted by academic & government administrations.

People who do science professionally know all about the competitive nature of the academic & research job market & "publish or perish". Just like anyone else, scientists are at least somewhat motivated by financial stability and career progression.

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

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

I think emulating existing darknet market infrastructure makes sense to address a couple of the issues you mentioned- like public reviews & messaging for employment + moderation & escrow/multisig for transactions.

+ your third point, I think another issue is basically verification: what's stopping cheap, non-refugee laborers from scooping job postings from refugees? I don't think it would be possible to verify the refugee status of someone without leaking identifying info.

I guess your last suggestion sounds like the best thing to do.

There is a lot of fundamental work on random projections [0] from Dasgupta et al., Achlioptas et al., etc.

For example, while random projections produced by sampling normal random variables like in the article are sufficient, there alternative approaches that produce transformations with "better" properties for big data- e.g. transformations with a sparse matrix representations.

For the basic background see the JL Lemma [1]- it turns with high probability, random linear transformations that map from high to low-d Euclidean spaces preserve inner products. In the context of compressive sensing, it's very much related the restricted isometry property [2].

Recently, it's been shown that random projections also appear in nature- the olfactory mechanisms of certain insects behave very similarly to random projections [3].

[0] https://en.wikipedia.org/wiki/Random_projection

[1] https://en.wikipedia.org/wiki/Johnson%E2%80%93Lindenstrauss_...

[2] https://en.wikipedia.org/wiki/Restricted_isometry_property

[3] https://science.sciencemag.org/content/358/6364/793/tab-figu...

This is basically adversarial training, which is a typical (& very practical) benchmark heuristic defense for this problem. An ongoing question is to precisely characterize when and how AT works. The line of work has also proved to be very fruitful for the theoretical community & has produced very general results about problems which can be solved by neural networks, but not other techniques- e.g. kernel methods.

https://arxiv.org/abs/2001.04413

It turns out that a similar technique, where you basically apply noise multiple times to a single image, and average predictions over all noisy images- equivalent to convolving your nn with Gaussian noise yields near state of the art bounds on provable robustness (under a specific class of attacks). The issue is the magnitude of noise you need in order to get practically robust networks is quite large relative to the data you are dealing with.

https://arxiv.org/abs/1902.02918

https://arxiv.org/abs/1906.04584

fyi the founder Igor Carron runs a pretty nice academic blog on compressive sensing (a bit less academic in recent months)

https://nuit-blanche.blogspot.com/

Also hosts the advanced matrix factorization jungle website which is a nice browse if you're interested in mf techniques.

https://sites.google.com/site/igorcarron2/matrixfactorizatio...

Someone correct me, but I think hMetis is one popular software/algorithm for multi-level graph/hypergraph partitioning. There is also KaHyPar which is a bit more academic.

Practical edge crossing minimization for large graphs (e.g. a billion nodes) is a really hard problem. An alternative formulations that's seen some use in graph drawing & circuit design software is to describe the embedding problem as an optimization problem: find an assignment of coordinates to nodes such that the edges (line segments) don't cross.

It turns out that estimating a lower-bound on the # of edge crossings of an embedded graph is equivalent to solving an SVM-type problem, and one can characterize the optimal layout as the solution to a nonlinear optimization problem. What's typically done, though, is to alternate between finding the bound for a fixed embedding & optimizing the embedding with something like gradient descent.

It gets a bit more confusing once you realize ic netlists are hypergraphs, and edges correspond to sets of nodes, (e.g. the "embedding" of an edge of a set of nodes is sometimes modeled as a rectilinear steiner tree).

For edge crossing/congestion minimization, the authors are Shabeer "Edge crossing minimization", Spindler "Congestion driven placement/RUDY", RePlace for a sota academic force-directed placement algorithm.

Relevant paper from Misha Belkin's group https://arxiv.org/abs/1909.12362

I wanna agree with this. I've seen a bunch of those animations before..e.g. https://www.deviantart.com/kirokaze

and it would probably be great to give the original artists a shout out.

Not necessarily for or against the work, but it's not just him working on this project solo like some mad scientist. There are a couple of other members on his team with academic publication histories as well.

I wasn't talking about methods for matrix multiplication.

The algorithms that are used by most ad recommendation services are deep enough themselves.

The ICML test of time award this year was for an adaptation of a technique used to study online decision making algorithms.

Some tier-A mathematicians working on some pretty fucking cool problems in online decision making and machine learning, or someone who writes shit like:

"Young people these days, I am told, are illiterate and cannot understand the written word. They can learn only from video. So should I speak to them in the language they understand?

Yes, I probably am a genius—but the issue is meaningless."

To add on: Zeyuan Allen-Zhu and Yuanzhi Li had a great theoretical paper earlier last year proving a fundamental discrepancy between the concept classes learnable by kernel methods and deep learning models.

https://arxiv.org/abs/2001.04413

One cool solution the wikimedia research team has been working on is to apply multilingual nlp. Like, how can someone detect a discrepancy in information presented in an article written in english & the same article written in russian?