Many Machine Learning Studies Don't Actually Show Anything Meaningful, But They Spread Fear, Uncertainty, and Doubt (theoutline.com)

Michael Byrne, writing for the Outline: Here's what you need to know about every way-cool and-or way-creepy machine learning study that has ever been or will ever be published: Anything that can be represented in some fashion by patterns within data -- any abstract-able thing that exists in the objective world, from online restaurant reviews to geopolitics -- can be "predicted" by machine learning models given sufficient historical data. At the heart of nearly every foaming news article starting with the words "AI knows ..." is some machine learning paper exploiting this basic realization. "AI knows if you have skin cancer." "AI beats doctors at predicting heart attacks." "AI predicts future crime." "AI knows how many calories are in that cookie." There is no real magic behind these findings. The findings themselves are often taken as profound simply for having way-cool concepts like deep learning and artificial intelligence and neural networks attached to them, rather than because they are offering some great insight or utility -- which most of the time, they are not.

Comments

Didn't we all assume that?

[nysus]

When AI can teach itself how to use a programming language from documents found on the internet or solve a long unsolved mathematical puzzle, that'll be something to talk about.

Re:Didn't we all assume that?

[gweihir]

This is not strong AI we are talking about here, it is weak AI. Strong AI does not exist. Weak AI cannot do anything that requires insight or understanding. It just can do statistical classification.

Re:Didn't we all assume that?

[HiThere]

Please demonstrate that a human can do something that requires actual insight as opposed to statistical calculation. Now prove that this wasn't done via statistical calculation.

The real problem that most AIs have is lack of grounding and a weak goal structure. But if they have decent grounding and decent ability to manipulate their environment, then you'd better pray that you got the goal structure correct, whether or not they are "strong AI". Cockroaches aren't strong AI, but just try to get rid of them. And the AI will make itself useful to some powerful group of people. (Possibly the group that caused it to be created, but that depends on the goal structure.)

Re:Didn't we all assume that?

[ceoyoyo]

"But just to give you an intuition: Weak AI cannot plan, cannot judge, cannot explore"

Some machine learning control systems are explicitly designed to form and execute plans. The ones that are designed to be trained in the real world are usually also designed to learn to do internal simulations (imagining what will happen if I do X) because feedback in the real world is slow.

Judging is pretty much what machine learning systems do.

Many reinforcement learning methods have a hyperparameter to explicitly control the amount of exploration the system does.

Are you sure you're not thinking of 1960's era tech?

It really is like human intelligence.

[hey!]

The human brain sees pattern everywhere it looks too.

I'm retired now but I've been doing a lot of reading and experimentation with decision tree based classification methods. I like these because the produce models you can examine critically, as opposed to so-called "deep learning" algorithms which produce results that you pretty much have to judge by their giving you the result you expect. It's not that that isn't useful in some cases, but I don't find it as interesting.

Re:It really is like human intelligence.

[swillden]

The human brain sees pattern everywhere it looks too.

Yep. Pattern identification system identifies patterns, news at 11.

OTOH, the ones I find interesting are the cases where ML identifies patterns that humans might not be able to identify. Sometimes this is less interesting for the potential to use a machine to identify these patterns than for the indication that patterns exist where we might think they don't.

The recent paper on ML "gaydar", where researchers trained a machine to identify sexual orientation from dating site photos is a potentially-fascinating example. In that one I suspect the algorithm was mostly keying on things like hairstyle and other indicators that the people in the photos might be deliberately using to advertise their orientation (this was a dating site), but there seems to be some evidence that facial structure also played a significant role. Of course, given that the most common genders and orientations (cisgendered heterosexuals) are clearly strongly correlated with certain characteristic facial structures, it's certainly reasonable to expect that the same genetic and development processes that determine gender and orientation and clearly affect face structure in the common cases should also affect face structure in the less common cases. But humans can't really see these patterns. Absent behavioral clues, people have very bad "gaydar". But that doesn't mean the patterns aren't there, just that we can't see them. If ML can see strong correlations between facial structure and orientation (and I don't think this one study proves that), that tells us something about the nature of sexual orientation and the degree to which it is expressed in the body.

Re:It really is like human intelligence.

[tomhath]

On one level, this study showed that you can correlate a conclusion (gay/not gay) with some pattern in the data. The problem is that it has only recognized a pattern in that particular data set.

The article mentions the factors were probably grooming and how the person posed for the picture. Okay, that might be somewhat useful if you are trying to guess sexual preference from a picture on a dating site. But not necessarily, because they threw out sample pictures that didn't provide the clues they were looking for, and (I suspect) selected test pictures that did have the clues.

Re:It really is like human intelligence.

[hey!]

The usual methodology for training is you start with a big sample of data and you randomly divide the data into records into two subsets; the first you use to train the model and the second you use to test the results of the training.

If there is no statistical difference between your training and testing groups, a better-than-random performance on the test data indicates that your algorithm actually learned something about the original universe of data. At that point you have the same problem you always have in statistics when you try to use your results: is some set of data you encounter in the wild so to speak really comparable to the data you build the model on?

One of the advantages of regression learning is that a classification your model produces is rebuttable. This is very important in a world where some courts are using proprietary software in sentencing to classify people by how likely they are to re-offend.

Re:It really is like human intelligence.

[laurencetux]

Good luck sorting out when the pattern is being matched wrong

even out of costume i would bet Male (hetro) Ballet Dancers might be miss-classed and a bit of thought could come up with cases where all four? possible options could be chosen wrong ( i think the set is MG ,MS , FG and FS)

Re:It really is like human intelligence.

[hey!]

A working knowledge of probability and statistics is probably more important than a working knowledge of calculus.

Many, many years ago when I was a student I took the famously difficult stochastic processes course at MIT. Recently I went through MIT Open Courseware's 18.05 lectures, and was amazed by how much the teaching of probability and statistics has changed... and for the better.

Trash science

[Hentes]

In the past it used to be stuff like "Scientists find that painting your room yellow leads to cancer!", now it's the same with AI. Turns out flipping through large amounts of statistical data until you happen upon a correlation is easily automated, and trash scientists will soon have to worry about their jobs.

And when it notices patterns..

[maitai]

AI notices patterns that detect cancer. Woot! AI notices patterns that determines crime rates amongst certain population groups! Fuq no. (I can't wait until someone calls AI sexist for it noticing that females get pregnant much more often than males)

Re:The AI's are coming! The AI's are coming!

[admin7087]

And it has been evolving at a steady pace since then. First, expert systems, then powerful defeasible reasoning systems and diagnosis, text-to-speech, rudimentary speech recognition, recommender systems. Now complex image recognition, music synthesis, intelligent search, working speech recognition and many more machine-learning based pattern recognition tasks. Given how young A.I. research is as a discipline, it has really made tremendous advances. Your phone can solve AI tasks that would have been considered inconceivable in the 60s. It may not yet seem human-like enough for you to be classified as intelligent behavior, but there is a threshold that will be passed probably still during this century.

Studies are not the problem.

[Gravis Zero]

Studies simply exist to inform others of a topic of interest. The problem is not the studies being scary, the problem is that highly technical information is inappropriately being repackaged and pushed out to the general public who have no insight into topic. The problem isn't the message itself, it's the messengers.