In the 21st century, statistics has had the odd distinction of being thrust into the spotlight in a way that it never has before in its centuries of existence. Before now, practitioners have mostly been on an island...or more like several islands that are multi-day journeys from each other by rowboat. It's created weird terminology, and even the initiated don't all use the same jargon. I have a degree in statistics, and one thing I learned in school is that if you pick up a textbook the first thing you have to do is figure out its internal terminology. Only the basest concepts or ones named after people tended to use the same names everywhere.
I think this is why computer science has been so successful at co-opting a lot of statistic's thunder. It's reorganizing a lot of concepts, and because everyone is interested in the results (image generation, computer vision, etc) it's getting a lot of adoption.
Amusingly, I thought that the blurb on MCMC the author quoted was pretty clear. That doesn't happen to me often.
> I think this is why computer science has been so successful at co-opting a lot of statistic's thunder. It's reorganizing a lot of concepts, and because everyone is interested in the results (image generation, computer vision, etc) it's getting a lot of adoption.
I think it's also that "computer science" can be perceived as having succeeded where statistics failed. And the terminology is frankly a lot more appealing: would you rather "train an AI algorithm", or "fit a model"?
Even the recasting of "model" as "algorithm" has a marketing benefit: models have uncertainty and uncertainty is scary, whereas algorithms are perceived as precise and correct.
Where are you viewing statistics as having failed, precisely?
Lots of well-documented success in traditional manufacturing and logistics type things. The "AI winter" wasn't a statistical thing at all - rather, stats have succeeded where logical programming had failed. Branded as "machine learning" or "data science," sure, but I don't recall a time of handwringing over stats failing in between.
(And I've heard "train a model" a thousand times more than "train an algorithm" from data folks in industry...)
> Where are you viewing statistics as having failed, precisely? <etc>
I don't think it's failed. But it certainly failed to capture the public imagination. Some people I talk to (fewer now than in the past) seem to think that statistics is old and irrelevant, and that machine learning is new and changing the world. Those people also don't usually have any idea of what they're talking about, but the idea is out there.
> "train a model"
But how often do mass media or industry publications talk about "models" as some kind of new hotness? It's all about "AI algorithms" now.
The terminology usage is especially surreal in the "digital insurance" space [0], where content marketing writers churn out breathless articles about how AI is the next big thing, as if insurance execs didn't understand what a model was!
It is my understanding that AI is being used in insurance not to supplant the risk models, but to automate human-driven tasks, such as reviewing claims or documents, or to augment risk models with new data.
That's correct. But the difference between an AI claim processing algorithm and the risk model lies more in how the model is used than in how it actually works.
Of course, the technical details of a text transformer differ significantly from the technical details of a decision tree or logistic regression. But when you zoom out a little, it's clear that many of the same core principles and techniques are used in both cases.
So the content marketers are right in that the next big thing is automating tasks like claim processing. But I find it at least a little bit silly because, at its core, an AI claim processing algorithm is not so conceptually different from a pricing model, and it's well within the realm of understanding for many people on the quantitative side of the insurance business.
I also want to avoid making particular claims or judgments about what AI "is". I will leave that to the futurists, philosophers, and AI researchers.
I think what the poster was implying was that most people create simple models, because they have a linear thought process, but they call it a model to sound sophisticated.
I was going for they call it a model to sound sophisticated - they call it training a model when they are really doing a linear regression.
Technically doing a pair of linest in excel and a goal seek to weight the output of each is Ai.
Sometimes the Groundbreaking Ai startup that just raised a gazillion dollars in their 10th up round, is just a plain old basic high school statistical model.
> in a way that it never has before in its centuries of existence
I'm not surprised by this at all. There has been an explosion in statistical research really since Fisher. While Bayesian Statistics was invented in the late 1700's, it wasn't really widely used until the 1950's (Fisher coined the term btw). The explosion in computing resources also caused an explosion in statistics, and especially in Bayesian Statistics. It isn't computing leveraging statistics but a symbiosis. Causal Statistics was pretty much non existent prior to the 21st century and major players like Judea Pearl didn't make their ground breaking works until then either.
It should be absolutely no surprise that statistics wasn't in the spotlight before. It was harder to do without computers. Obtaining vast amounts of data was substantially more difficult. Especially with respect to accuracy. Determining causal relationships was essentially a crap shoot.
But this is also a great example of why your theoretical work may not be useful now but can be the cornerstone of modern science a century later.
This. Statistics is a very young field. The entire idea it rests on[1] is highly unintuitive, so it's no wonder it takes time.
----
[1]: Exchangeability, the fact that you can make useful progress in analysing a case by ignoring most of the information about that specific case and instead lumping it into a reference class of "similar" cases, where similar ultimately rests on subjective judgment, other than in a few special cases.
I definitely agree that statistics is a young field, but I would not agree with your definition. This is naive and statistics is much more. As HN readers that know me will attest, I'll frequently rant about how low order approximations can frequently lead to inaccurate or even results that lead in the wrong direction. Statistics actually allows us to see this! There's many famous paradoxes that are really dependent upon this. For simplicity sake's I'll reference Simpson's[0] and Berkson's[1] which happen because of inappropriate aggregation (which inappropriate aggregation is one of the most common errors in statistical modeling, but not always easy to notice).
Rather what statistics is about is modeling things that either have imperfect information and/or are probabilistic in nature (yes, there are real probabilities). The point isn't to ignore information (we actually want as much as we can get, which is why we're in the information age and why statistics has exploded with innovation) but to recognize that with observations and sampling that we can extract patterns that still allow us to make predictions without all the information. For example, statistics is commonly used in thermodynamics because there is imperfect information[2], in that we can't measure where each individual particle is at a given point in time. Instead we can aggregate information about the particles and then make aggregated predictions about future states. It is important to note that the predictions are also probabilities.
[2] thermodynamics also involves true probabilities but I think the case is still motivating. We can also use statistics with purely deterministic processes with perfect information, but it is frequently computationally inefficient.
I agree with everything you say! But I would still argue that the things you describe (prediction of the partially known based on limited information) is built on probability theory (sort of applying it in reverse), which in turn is based on assigning cases to reference classes and treating them as exchangeable in the ways that matter for the analysis at hand.
To be fair, naming things is also one of the hard problems in computer science.
Graphs may have vertices or nodes. Traversals of a graph may be called paths and walks or simple paths and paths. The length of a path may be equal to the number of nodes or edges in it. The height of a tree may be equal to the depth of the deepest node, or it may be off by one in either direction.
In my subfield, some people talk about strings and substrings while others talk about words and factors. A Lempel–Ziv compressor may factorize the input or parse it into phrases. The Burrows–Wheeler transform, the FM-index, and the compressed suffix array may all refer to the same thing, or there may be 2 or 3 distinct meanings.
I don't think I've ever written a paper without naming at least one new concept and renaming at least one old concept.
I think this is why computer science has been so successful at co-opting a lot of statistic's thunder. It's reorganizing a lot of concepts, and because everyone is interested in the results (image generation, computer vision, etc) it's getting a lot of adoption.
Amusingly, I thought that the blurb on MCMC the author quoted was pretty clear. That doesn't happen to me often.