Merry Christmas! Our Walter Bradley Center director Robert J. Marks has been interviewing fellow computer nerds (our Brain Trust) Jonathan Bartlett and Eric Holloway about 12 overhyped AI concepts of the year. From AI Dirty Dozen 2020 Part II. Now here’s #8. Sick of paying for health care insurance? Guess what? AI Hype is better! Or maybe, wait…
8 In Our AI Hype CountDown: AI Is Better Than Doctors! “Is AI really better than physicians at diagnosis?” starts at 01:25 Here’s a partial transcript. Show Notes and Additional Resources follow, along with a link to the complete transcript.
Robert J. Marks: We’re told AI is going to replace lawyers and doctors and accountants and all sorts of people. So, let’s look at a case of the physicians. This was a piece on Mind Matters News. Eric, what do you think? Do you think that AI will ever be better than physicians at diagnosis?
Eric Holloway (pictured): Well, I don’t know if they ultimately will or will not, but right now they definitely are not. This particular author, he took a look at 10 years worth of studies for deep learning algorithms on medical problems, and only two of them actually relied on randomized trials while 81 were not randomized.
This means basically people can just pick and choose the type of data that makes their algorithm work well. So, really their results don’t really tell us anything about how well their stuff works in the real world.
Note: Eric Holloway is referring to the problem that, if researchers know what they are looking for, they tend to find it. That applies to using AI as much as to anything else. The only practical solution is randomization. Picture a Bingo! game. The researcher does not know what numbers to expect:
“What’s the difference? Randomized tests make data harder to manipulate because the researchers don’t know what data they will be assigned. That makes data harder to manipulate, consciously or otherwise, to favor an outcome.
Even better, double-blind tests—where, for example, neither the doctor nor the patient know which drug is real and which is a placebo—further reduce the likelihood of bias helping determine the result.
These are the standards we should apply to all AI-based medical devices, even if they simply assist medical personnel. Otherwise, without anyone meaning to be dishonest, the fox is guarding the henhouse.” Brendan Dixon, “Is AI really better than physicians at diagnosis?” at Mind Matters News
Robert J. Marks: Gary Smith, who is also one of the fellows of the Bradley Center, talks about the idea that when you publish a paper based on statistics, you got a problem.
Note: The problem is greatly increased by Big Data, which—as Gary Smith explains—features many meaningless patterns (he calls them “phantom patterns”) that occur naturally, simply on account of the size of the file.
Randomness is like a bad road. It is bumpy, not smooth, but it is still random: “We do not fully appreciate the fact that even random data contain patterns. Thus the patterns that AI Hype algorithms discover may well be meaningless. Our seduction by patterns underlies the publication of nonsense in good peer-reviewed journals.”
Over time, the phantom patterns will fade and disappear—but often not before they are published in peer-reviewed journals.
Robert J. Marks (pictured): Gary points to the Texas Sharpshooter Fallacy: If you have a barn door and you paint a bunch of targets on it and you shoot at the barn door with an arrow, you’re going to get close to a bullseye if you have a thousand targets up there.
So, there was this one case about pancreatic cancer, and they began to look at correlations with pancreatic cancer. And, well, they thought it was caused maybe by smoking. No, it wasn’t caused by smoking.
What about, I don’t know, what about chewing tobacco? No. Chewing tobacco. Drinking tea? No. How about smoking cigars or pipes? No, it didn’t correlate. What about drinking coffee? Oh my gosh, there was an incredible correlation there.
So, they publish this in the New England Journal of Medicine and coffee futures fell and people stopped drinking coffee. And, in fact, in the end it turned out that it was totally just a coincidental correlation.
And, subsequent studies showed that the correlation was just coincidental. In fact, another study said if you drank a lot of coffee, your chances of contracting pancreatic cancer were improved. So, it’s just crazy. And, I think that that’s one of the problems that we have. But, you hold out promise for the future maybe, huh?
Note: From the New York Times 1981: “Dr. Thomas C. Chalmers, president of the Mount Sinai Medical Center and dean of its medical school, commented that the investigators who questioned patients on their prehospitalization coffee habits knew in advance which ones had cancer. This could have introduced unintentional bias in the results, Dr. Chalmers asserted.” Not exactly randomized.
Eric Holloway: Yeah. I would say probably if you restrict the domain enough, you’re going to be able to pull out some stuff. But, the other problem, too, is how they tend to build these systems. They get a data set from some doctors, and then they just go off for a bunch of years and try to make some algorithm that scores highly.
What they really need to be doing is working much more closely hand-in-hand with the doctors and trying to optimize particular parts of their workflow with these algorithms instead of just trying to replace the doctors.
In our countdown for the Top Twelve AI Hypes of 2020…
9: Erica the Robot stars in a film. But really, does she? This is just going to be a fancier Muppets movie, Eric Holloway predicts, with a bit more electronics. Often, making the robot sound like a real person is just an underpaid engineer in the back, running the algorithm a couple of times on new data sets. .
10: Big AI claims fail to work outside lab. A recent article in Scientific American makes clear that grand claims are often not followed up with great achievements.
This problem in artificial intelligence research goes back to the 1950s and is based on refusal to grapple with built-in fundamental limits.
11: A lot of AI is as transparent as your fridge A great deal of high tech today is owned by corporations. Lack of transparency means that people trained in computer science are often not in a position to evaluate what the technology is and isn’t doing.
12: AI Hype is going to solve all our problems soon! While the AI industry is making real progress, so, inevitably, is hype. For example, machines that work in the lab often flunk real settings.
Originally published at Mind matters