My favorite academic paper ever [0] was a comparison against a bunch of dimensionality reduction algorithms and 100 year old PCA was tough to beat!
Glad I was able to pivot my career out of AI and ML. My PhD wasn't at Stanford, MIT, et al so I couldn't find any jobs doing the "actual research" - if they existed at all outside academia.
EDIT to add another funny "frustration" paper more directly related to ML [1]. I consider DR is more of a data analysis thing.
[0]: van der Maaten, et al. Dimensionality Reduction: A Comparative Review https://members.loria.fr/moberger/Enseignement/AVR/Exposes/T...
[1]: Dacrema, et al. Are We Really Making Much Progress? A Worrying Analysis of Recent Neural Recommendation Approaches https://arxiv.org/pdf/1907.06902.pdf
I wish AI/ML was added to Pure Data [0] or Max [1]. This would require all your skills if you could help out :)
[0] https://puredata.info/ [1] https://cycling74.com/products/max
For some reason, in the computer world, scientist, engineer, and technician have been merged into a single role. Full-stack, if you will. But as the success of the separation of roles in other domains shows, it does not have to be like this.
I mean in software the theory is so close to the practice that there's little purpose to create a role of software technician for the software engineer.
You could say developer or programmer means technician, but it's not really the case in most companies, because developers do engineering work, too.
A lot of projects don't even have a proper spec, because the program can serve as a de facto spec. In a company where you need a working program, why would you cause translation errors to slow you down if you don't need to?
Separating spec and implementation is possible, but it's an inefficiency that needs a good rationale in a profit driven business. Which is why you see proper specs with later implementation only there where these costs are deemed appropriate. Like safety critical systems in aerospace and automotive.
It does not have to be like this, but it's an efficiency optimization that's just not possible in other engineering domains.
In classical engineering the closest we've come to this is rapid prototyping with CAD tools that allow simulation. Because it's much more efficient when the engineer can do the implementation and find design errors before they're mass produced.
Nitpick: technicians and technologists are distinct in Canada. They operate at different levels in the stack as shown in [1].
[2] has some requirements for becoming certified as a technologist or a technician.
I agree with the rest of what you’re saying. However, this system has limitations (moving up in levels) that should be improved on before this model is adopted or imitated.
[1] https://asttbc.org/wp-content/uploads/2019/02/Level-of-Work-...
However, most of the people who want to do the job want to do it because they heard Google pays six-figure salaries for it. But Google pays six-figure salaries for the job because it takes a PhD to do it right. When the job becomes a job that everyone can do, Google won't pay six-figure salaries for it any more - to PhDs, or anyone else.
At that point, once more, you'll need a PhD to do a job that Google pays a six-figure salary for.
The goose that lays the golden eggs gets killed over and over again. Some people get in early and get a golden egg. The rest are left to suck ordinary eggs and accuse the others of cheating.
But, who was it who killed the goose in the first place?
Some of us are actually in it to scratch an intellectual itch, not for the money. I gave up a lucrative career as a software dev to study the subject I'm really interested in and took a huge pay cut for the privilege of working in the only environment that offers anything near the freedom to research whatever satisfies the little gray cells. The OP's turn of phrase about gatekepping sounds like a bitter joke to me. My field gets maybe one or two new entrants a year, folks like me who don't know what's good for them and can't tame their intellectual curiosity enough to get a real job. Most PhD students, most researchers of any level, don't want to gatekeep anything, we want to tell the world about our work. But most people don't give a shit, unless Google is interested. Unless you can grab my research and be an instant millionaire nobody cares.
A bug species detector that works almost all the time will get commoditized. A bug species detector that fails gracefully won’t get commoditized until later.
An insurance risk predictor will get commoditized. An insurance risk predictor that you’re confident is fair and unbiased won’t get commoditized until later.
A chat bot that gives you customers useful information about your topic will be commoditized. A chat bot that also never will say anything to tarnish your brand (e.g. sexist/racist/wrong) won’t be commoditized until later.
A last example illustrates it perfectly. The majority of automating driving was solved pretty quickly. That last mile of reliability has taken forever.
I’m convinced that knowing what’s going on under the hood will still be valuable for a while yet, because we are just not starting to really face ML failure modes at large scale.
A smaller set of people than those using them.
When talking to the PhD boss of the gatekeeper group, he told me that in practice a lot of practical ML work is done with a handful of algorithms and most of what he learned on his PhD is not useful in the industry. And a lot of complexity now sits behind tools (what TFA says).
He was suggesting that, in most applications, a ML expert could be useful as a consultant for a brief period early in the project, but 1) the amateurs wouldn't do too bad by picking algorithms themselves and using the tools for tuning after some study of the fundamentals and 2) in case an expert is consulted, the rest of the team can just run with their suggestions for 90%+ of the duration of the project.
Unsurprisingly extraordinary results in AI require extraordinary commitment. It may or may not make sense from a business perspective.
...who won't do as well as people who understand the business domain and that "good enough" isn't that hard to achieve with some pretty elementary stuff (regression, xgboost).
PhD's have been trying to act as gatekeepers of "Data Science" for the past decade. It's only getting easier for people to apply this stuff. Unless you're doing actual research in these algorithms, there is little need to "understand the math" beyond an undergraduate level.