I think now really is different. Part of that is algorithmic advances like deep learning, as shown in this Nature paper.
An even larger part of it is that the financial incentives are flipping due to value-based care. In 1979, a hospital that implemented an expert system for accurate diagnosis may, paradoxically, see its revenue fall. Nowadays, with ACOs, risk-based contracting, and bundled payments, the financial incentives create tailwinds rather than headwinds for large-scale adoption of AI in medicine.
Contrary to popular belief, the medical system can absorb new techniques very quickly--when incentives are aligned. And they are now becoming aligned.
Many of the barriers standing in the way of wide-spread use diagnostic software are not technological in nature.
Ignore those barriers at your own risk.
The deep neural net visual diagnostics are different. They are learned on pure pixels as we would from photons striking our retina, signals traveling all the way back to our visual cortex for learning. There is no assembling of thousands of rules here and therefore the system is less brittle.
These new systems get more powerful with more data given to them. Expert systems required humans to craft rules from the data, and therefore required constant maintenance and is fallible to human error.
So yes, some of the barriers are definitely technological in nature.
IIRC, MYCIN had several hundred rules. The researchers in this article had to pre-process 130,000 labeled examples. If you see a misclassification in an expert system you can at least backtrack and identify the individual rules that contributed to the failure. AFAIK, systemic errors in training data are much more difficult to detect and fix.
I think people tend to overstate the practical issues with expert systems and understate the issues with deep learning, partly because we have decades of experience with real-life deployments of the former and relatively little experience with the latter.
Yes, the explainability of expert systems is the only thing going for it.
Yes, there are issues with both, but we are really debating different solutions for different problems. For visual recognition, there is no doubt in my mind that deep learning is king.
It would be interesting to compare that with the current state of the art in the field, and see if ML can contribute new scientific/medical theory as well.
This is most simply because, whatever the algorithm is trained to do, it's certainly trained better to do that thing than to introspect. Introspection is a separate skill!
But there's also a more insidious element: introspection (in humans, at least) tends to result in the creation of a lot of "personal concepts" that don't map to well-known common concepts. An introspection on one mind must necessarily result in a taxonomy that contains terms for the tiny, unique features that only that mind has—which makes it very, very hard to communicate one's personal introspections to others. (You might call this a kind of overfitting: the introspection capability becomes optimized for that one mind, but ceases to translate well to features in other minds—like human minds.)
I'd place a much stronger bet on our ability to train one AI to "stare at the brain" of other AIs as they make decisions [tons of them, as its training data], with the expected output being a general theory on common AI features responsible for the given calculation step. A computer psychologist, of sorts. :)
Of course, you could include such a pre-trained model as a "module" alongside the AI itself, and call the combined system "one AI" if you like.
Indeed it is. It's something that a second algorithm (perhaps ML, perhaps not) would do.
And this is beginning to remind me of Society of Mind.
It's a neat approach that I've used with some random forest classifiers.