groceryheist

Here's a non-paywalled version. https://ndg.asc.upenn.edu/wp-content/uploads/2016/04/Centola...

This is just an example of a well-known recent sociology paper that can appeal to science-minded people unfamiliar with sociology.

By the way, it's a danger to rely too much on mathematical formalism as a heuristic for deciding whether a research field is scientific or valid. String theory for example lacks empirical evidence (and maybe even the possibility of empirical validation altogether).

How about https://www.journals.uchicago.edu/doi/abs/10.1086/521848

?

This is a very narrow view of potential applications of facial recognition. I take it for granted that this technology has enormous affordances for social control. "Criminality" may be an out-of date notion of the kinds of behaviors that governments and other institutions like those in workplaces (e.g. the academy, employers, corporations). Surveillance has major political implications and we should not treat the advancement of technology as independent from the society and the political economy.

Surveillance is the key concept in a major design of social control, Bentham's "Panopticon." When anyone may be watched, people must live expecting that they will be watched, and the will of those with the power to create sanctions will enacted. This will be a historical novelty if facial recognition systems become widespread. Historically, only limited public venues were subject to effective widespread surveillance (except in extreme authoritarian cases like East Germany). Increasingly, online spaces are subject to surveillance that is limited by browser instrumentation, with facial recognition it will be possible to track people offline as well, if you have access to the cameras.

Surveillance is the twin of transparency, which can be thought of as possible tool for democratic control. I do believe that transparency i.e. through freedom of information acts has some positive characteristics for democracy but has a weak empirical record compared to panoptic control. Facial recognition might have affordances for collective action, such as making it easy to produce evidence of excesses or corruption.

That said, maybe we should not give the authorities the power to implement technological changes without the consent of the governed? Technology moves faster than policies can adapt to them, but old laws and institutions like the police evolved in times when the tools of day could not have been imagined.

Edit: Regulating technologies like facial recognition popularizes the control of social systems which counter acts authoritarianism and high concentration of power and resources.

This is hilarious and previous HN commenters are in sore need of political reeducation to free them from the bourgeois counter-revolutionary ideology clouding their senses of humor.

Totally agree. This is maybe the one thing that could potentially bring me back into the Microsoft ecosystem.

Seems pertinent to discussions of Facebook's new cryptocurrency pseudo-bank no?

I think it might just be like a psychological thing to try to prevent mistakes. Like a dialogue box asking "Sure?" before you delete your harddrive.

To answer your question:

ggplot2 : plotnine is quite good ggplot2 clone based on matplotlib. I feel like ggplot2 is a bit better and more complete, but if you want to do something that isn't supported it's harder for me to hack than matplotlib.

Tidyverse: To me, ggplot2 is the only essential part of the tidyverse. Lubridate is also good. Most others seem like semantics and syntax sugar. I prefer data.table, which is similar to Pandas. DT is super fast but imho Pandas has a more intuitive and consistent API (and if you want a speed up for large N then dask might work).

I use both R and Python on a regular basis. I choose Python for lower-level stuff, automation, parallelism / concurrency, and R for bespoke statistics. I use both for everyday statistics and plotting, but I feel that R has light advantages. I feel like if you're comfortable switching languages there are good reasons to use both. It's also important for me because I work with different teams that have different practices and preferences.

Getting a definition or spelling correction for a word or a list of synonyms and antonyms, maybe a brief etymology, without interrupting my workflow at the command line or opening a slow heavy program like a web browser or office suite.

And what was wrong with dict?

> Distribute binaries created in whole or in part from the source code.

How exactly is that different from what's happening here?

I am a little worried about the accuracy of sentiment classifiers, relative to how often I see them used and in applied research settings. They work well for binary classification on Tweets and IMDB reviews and not much else (I guess they do alright at 93% if you magically have a parse tree).

Thanks for this effortful comment! I really appreciated your point about how the US system is good for those who push themselves. It makes me curious to consider the myth that the US system produces "creative" individuals in a way that China supposedly fails to do.

I agree that the US's academic and intellectual culture is good for the intrinsically motivated, but, speaking as a graduate student who teaches classes at a top public US university, most undergraduates arrive thoroughly institutionalized and are primarily motivated by their grades. I believe in project-based learning that gives students an opportunity to apply their skills and knowledge to problems they are personally interested in, but it is difficult to do this effectively when students have been trained to be motivated by their grades. I am not trying to focus here on criticizing the grade-based system (with which I have many qualms), but the broader culture that we see in early education. I think your point about teachers offering concrete feedback to students in terms of study skills and dedication is important. Much of the US middle class seems to expect success without feeling that they have to earn it, even as they expect others with less privilege to work harder. The US is experiencing the downside of an individualistic culture: those in secure socio-economic positions do not feel the need to work hard in order to help provide to the collective good.

Mother of all clickbait titles.

I read the first few pages and didn't see any content about R. The first chapter is an introduction to scheme. I scanned it and didn't see any R code all. It's an introduction to scheme that builds up to K means.

R does suck in so many ways, especially the object oriented systems and the horrible default libraries with so many inconsistent idioms, but we put up with it because mainly because it provides the most advanced open source ecosystem for modeling and visualizing data and it is also pretty fast, especially with data.table. If being fast is important we have Julia as a competitor. If having a better designed language with a reasonably complete ecosystem and sometimes being slower is OK we have python. Like pretty much every other domain area, there just isn't a competitive niche in data science for a lisp.

Also, R is already pretty functional and Lisp-like, sure it doesn't have prefix syntax, but it does have first class functions. Using various types of map (lapply, apply &c) is practically mandatory after a certain point.

I love the Wizard Book too, but this is embarrassing.

There are also female computer nerds that are not conventionally attractive. It might be worth trying to make friends with someone less intimidating.

You upvote/downvote results. Crowdsourcing search ranking, what could go wrong?

Edit: typo

Good. We don't need more monopolistic conglomerates.

| does anyone know if more recent machine learning techniques have made their way into the field?

Oh god I hope not.

While behavioral genetics is claiming victory in the replication crisis, it should not rest on it's laurels. Replicability is a sign that studies can be conducted consistently. However in the case of behavioral genetics, one fears that all they have managed to do is reproduce and scale up erroneous research designs.

Consider finding 3: "Heritability is caused by many genes of small effect"

This is exactly what you would expect to find if you do a regression with 1000s of variables: many effects, mostly small ones.

Of the 10 findings only 1 is clearly not a fact about statistics in settings where dimensionality is very high relative to sample size.

Finding 5. The heritability of intelligence increases throughout development

Observe that this finding is based not on associations between particular genes and traits, but instead on twin studies. Twin studies have their own problems when it comes to causal identification.

That said, the overall claims here make sense and have some empirical support. My point is that the field has long made outsized claims based on dubious science that is legitimized by it's appropriation of genetics (even though it is a branch of psychology, not biology).

Panofsky's book "Misbehaving Science: Controversy and the Development of Behavior Genetics" is excellent.

The ideas from this article are really cool, and the design is beautiful. I see these techniques as providing the ability to partially interpret models. While clearly useful to practitioners seeking an intuition for what their models learn, it appears we are still very from the ability to thoroughly audit deep learning computer vision models.

I wonder if in the long run, making models that are both effective and interpretable can be done by first building a black box model, and then interpreting as much as possible it using clever ideas like those from the article. The interpretations of the black box model can inform the design of a relatively simple bespoke model. The bespoke model may never outperform the black box at prediction tasks, but in many applications the ability to perform audits and estimate of uncertainty should be worth it.

Newsblur works great for me. It has some slick machine learning features to help manage complicated feeds.

Yes, multimodality is often a problem for mcmc clustering or dimensionality reduction. However, if you use the SVD method to estimate PCA you only have a bimodal distribution since SVD is identified up to the sign. Asymmetric initialization is usually enough to solve the problem.

This thread has some good examples of PCA implementations in STAN. https://groups.google.com/forum/#!topic/stan-users/5R2-QUDiy...

As someone who just implemented hierarchical probabilistic PCA in stan, I agree that it takes finesse, but it is no means impossible. Doing this sort of work efficiently in stan seems to require a some degree of understanding about how the sampler works. It also may require really thinking through your model. It saves you from deriving your own conditional distributions and writing a gibbs sampler, but you're going to have to do some analysis if you want to fit models of certain complexity.

KL divergence minimization (variational inference) is typically a weak approximation to the model you specified. I have seen it produce inferences on simulated which are just plain wrong. These "wrong" models are still often good predictors, so whether variational inference will work well for you depends on whether you care about making valid inferences or just doing prediction.

Any reason to use Scala native when Rust exists?

The "margin of error" reported in opinion polls is typically just a function of the sample size and sample variance. The error is assumed to be due to statistical variation, not to problems with measurement. Measurement error is a bigger problem than statistical variation because statistical variation can often be assumed to be unbiased.

My comment above is not about statistical variation, it is about measurement error. Inaccuracies in models for labeling images are a form of measurement error. Having humans label a (often stratified) sample of images is required to understand the measurement error. This requirement is the main difficulty when it comes to generalizing learning from one context to another. If you use multiple independent humans to label all your images, you can understand the errors made by the humans very well.

Hackernews likes this as a practicable application of transfer learning. Fans of machine learning want to see transfer learning as more than a cool trick.

Unfortunately, there is really no good reason for someone seriously interested in accurate information, like a researcher or journalist, to use machine learning for this particular task. Labeling a couple thousand images yourself or with friends is not that big of a task. Do it over a few evenings while watching TV and drinking beer. You could have mechanical turk workers do it for you for a few hundred dollars. In either case you will get extremely reliable information. If you use multiple judges you will have a good estimate of uncertainty for every classification. There is no way transfer learning can provide this uncertainty information.

The main advantage of this technique remains the ability to quickly label very large amounts of data on the order of hundreds of thousands of rows, or thousands of columns. For smaller data, machine learning can sometimes achieve marginal improvements in predictive performance through model complexity. However prediction in smaller data regimes is mainly useful for out-of-sample prediction. The machine learning paradigm offers limited support for measuring uncertainty for out of sample predictions, which is super important if you are a researcher.

One capability of transfer learning could be to support many many applications from one model, but I have yet to see demonstrations of this in practice. The problem is that knowing how well learning has transfered requires measuring generalizabity and so cannot be done blindly.

Don't represent 1-dimensional quantities like proportions or numbers of people as area. This is actively misleading because area it squares the magnitude of differences.

Why do all these features of a colony need to be part of the same thing?

I see 1. Reputation 2. Governance based on reputation 3. Division of labor 4. Team membership 5. Organizational ownership 6. Dispute resolution

It is not obvious what this particular set of organizational features is necessary or sufficient for. Even worse there are likely many different ways of implementing them and it is not obvious how implementation details will affect resulting organizational dynamics. A better approach would be a modular organizational design with different plug-ins that satisfy different organizational requirements. This would allow organizations to choose their rules and to innovate instead of having to pick up a complex model of organization off of a shelf. A modular design would let you start small, build a minimum set of requirements to power some organizations and then scale up the complexity from there.

In my experience consulting in boring offices of companies that provide business reporting and compliance services a lot of the people only have a few hours of work to do most of the time. But a lot of the processes are manual and brittle and so every so often there is a crisis and if they didn't have everyone to minimize the impact of the crisis on their customers they would lose business fast.

I bet there are a lot of companies where productivity is low because people don't make extra work for themselves, do what they need to meet low expectations but the company needs to keep people around because they can't build processes that don't need the people.

It's like mit scratch. Except the underlying purpose is to sell junk food, the only sprite you can control is a corporate mascot, and you do so alone instead of in a social platform.

The focus on problem solving may be good for helping kids learn and practice solving problems procedurally. The free form nature of scratch gives a lot of freedom for goofing off, which can be a good thing.

http://scratch.mit.edu