IMO a part of the problem here is that you have a misunderstanding on the part of deep learning people. They look at radiology, and they say "these people are just interpreting these pictures, we can train a deep learning model to do that better".

Maybe there's a bit of arrogance too, this idea that deep learning can surpass human performance in every field with enough data. That may be the case, but not if you fundamentally misunderstood the problem that needs to be solved, and the data you need to solve radiology, for instance, isn't all in the image.

Somewhat related: another area where DL seems to fail is anything that requires causal reasoning. The progress in robotics, for instance, hasn't been all that great. People will use DL for perception, but so far, using deep reinforcement learning for control only makes sense for really simple problems such as balancing your robot. When it comes to actually controlling what the robot is going to do next at a high level, people still write rules as programming code.

In terms of radiology and causal reasoning, you could imagine that if you added extra information that allows the model to deduce "this can't be a cancerous tumor because we've performed this other test", you would want your software to make that diagnosis reliably. You can't have it misdiagnose when the tumor is on the right side of the ribcage 30% of the time because there wasn't enough training data where that other test was performed. Strange failure modes like that are unacceptable.

Expanding on this, particularly regarding causal reasoning and rules, what I find especially puzzling is the desire to apply deep learning even in cases where the rules are explicitly known already, and the actual challenge would have been to reliably automate the application of the known, explicitly available rules.

Such cases include for example the application of tax law: Yes, it is complex and maybe cannot be automated entirely. However, even today, computer programs handle a large percentage of the arising cases automatically in many governments, and these programs often already have automated mechanisms to delegate a certain percentage of (randomly chosen, maybe weighted according to certain criteria) cases to humans for manual assessment and quality checks, also a case of rule-based reasoning. Even fraud detection can likely be better automated by encoding and applying the rules that auditors already use to detect suspicious cases.

The issue today is that all these rules are hard-coded, and the programs need to be rewritten and redeployed every time the laws change.

There's a perception in the DL field that encoding things into rules is bad, and that symbolic AI as a whole is bad. Probably because of backlash following the failure of symbolic AI. IMO the ideal is somewhere in the middle. There are things you want neural networks for, and there are also things you probably want rules for. The big advantage of a rule-based system is that it's much more predictable and easier to make sense of.

It's going to be very hard to engineer robust automated systems if we have no way to introspect what's going on inside and everything comes down to the neural networks opinion and behavior on a large suite of individual tests.

> The issue today is that all these rules are hard-coded, and the programs need to be rewritten and redeployed every time the laws change.

The programs are probably not being rewritten from scratch. I would argue that: the laws are, or basically should be, unambiguous code, as much as possible. If they can't be effectively translated into code, that signals ambiguity, a potential bug.

I wasn't alive in the 70s, but it feels like there's a counter-bias against expert systems borne out of those failures.

"If you're putting in rules, you're don't know how to build models."

But that's probably the difference between people having success with "AI" and banging their heads against the wall: do what works for your use case!

I have once seen an AI tool to determine what needed to be reported.

I found this remarkable, as there were clear (yet complex) rules on what needed to be reported, otherwise even the regulator wouldn't know what it was supposed to check.