On March 30, 2020 12:11:03 AM EDT, Mario Carneiro <[email protected]> wrote:
>This approach sounds like "expert systems" (
>https://en.wikipedia.org/wiki/Expert_system) from 80's AI. My
>understanding
>is that it's more or less discredited these days

I think that is too strong a claim. Expert systems still have their uses. 
However, they require you to be able to define very clear and ambiguous rules. 
When humans can provide those rules, and the ruleset is not too large, expert 
systems can work very well.

So it is not that they are discredited. The problem is that they turn out to be 
a good but relatively niche solution for a relatively narrow set of problems. 
When they match the problem space, they are great, but most problems don't map 
well to them.

> and machine learning
>systems have little to no resemblance to expert systems.

That is true. A serious advantage of machine learning systems is that they can 
handle probabilistic results in a cleaner way than some alternatives, and as 
already noted, they can learn from data sets.

Most machine learning approaches are terrible at explainability. They can give 
you a prediction, but not a human understandable reason for why they make that 
prediction. There are a few approaches that provide explainable results, such 
as decision trees, but they have their own problems. I am skeptical that many 
would want to apply a machine learning system to medical diagnosis without also 
being able to provide an explanation.

And that is nothing to say of the fundamental problem, that so far they have 
not done very well at broad medical diagnosis. There have been some significant 
successes in very narrow applications, such as looking for tumors in images. So 
it's not fair to say that machine learning is useless for sny medical 
diagnosis. It's just that the current ML technology is still in its infancy 
when it comes to its application in medical technology. Medicine is hard, and 
while machine learning can sometimes do some impressive things, we are still 
learning what it is good and not so good at.

In contrast, I think the OpenAI work to create proofs is a far more amenable 
task for today's machine learning systems. That's not to say it is easy, it is 
not, but I suspect that is a far more attainable goal for machine learning as 
it currently exists. I'm very excited to see what OpenAI  publishes when they 
do.



--- David A.Wheeler

-- 
You received this message because you are subscribed to the Google Groups 
"Metamath" group.
To unsubscribe from this group and stop receiving emails from it, send an email 
to [email protected].
To view this discussion on the web visit 
https://groups.google.com/d/msgid/metamath/F7DFF56D-E50B-4CB0-80FA-8FF9E8028B4A%40dwheeler.com.

Reply via email to