For what it's worth, I completely agree with the sentiment presented here that we're far enough away that we don't know what direction even we need to start down to make the best progress, which in my view makes any sort of prognostication of questionable value. It's also why I think it's not that productive to have people like Stephen Hawking arguing about future AI as a threat that somehow needs to be mitigated. Even if he's right in the general point, we know so little about how that future system would look that it would be like asking Isaac Newton to make cars safer.
Even worse, we have some strong theoretical results suggesting that none of the techniques we have developed so far can ever lead to AGI, and that refers very much to machine learning primarily.
I'm talking about the work that was done on inductive inference, and then PAC-learning, in decades past (starting in the 1960's). There are many negative results and very few positive ones.
I wouldn't say I've reviewed the bibliography extensively (it's a huge area spanning a couple of fields), but from the little I've read I'm getting the feeling that, if you want to be able to learn any function in polynomial time, with a polynomial number of training examples, and with arbitrary error, you're basically stuck with the easiest classes of concept to learn. And if you want to learn specific distributions or representations (like DFAs or PCFGs) then you can just forget about it.
Which reminds me of something that Andrew Ng has said, about how recent results in machine learning and deep learning only reached the "low hanging fruit" and further progress is going to be much harder. No references to all this- I'll dig 'em up if required.
Any result which predicts that human intelligence cannot exist.. has to have an error somewhere. Perhaps the error here is in assuming that a learning algorithm able to fit any function in polynomial time is needed in the first place.
People are too ready to leap at results that say, "Well, machines can never do X, and I can do X, therefore human-level AI is impossible."
In general, my counter to that would be "you probably can't do X either". It's just that you can do it well enough. You don't encounter the weird cases where your ability fails. But if that's allowed as an argument, then of course machines are more powerful too. I can absolutely solve the halting problem if "solve" can be interpreted as "almost always come up with the correct answer for the kinds of problems people encounter in the real world".
Also, don't forget what Terry Pratchett said, about how thousand-to-one odds crop up one time out of ten. The problems with AI in the past started when people tried to attempt to solve anything beyond the trivial, toy-problems they could demonstrate were solvable "in the lab". Like SHRDLU, which solved blocks world, but could not really scale. The same goes with machine learning today: some problems are solvable in restricted settings, or in a lab environment, under controlled conditions and with safe assumptions etc, but once you try to do the same thing in the wild...
You can even look at the big success of Deep nets for examples of that: image processing and speech recognition. Whereas it's possible to get really good results in both fields, image processing still relies very much on having good quality images to work with, and speech recognition is really crap if you have a bit of an accent, or a speech impediment etc. (at least that's what I gather- these are not my specialties).
Dana Angluin, Inductive inference of formal languages from positive data. [http://www-personal.umich.edu/~yinw/papers/Angluin80.pdf]
Dana Angluin, Learning regular sets from queries and counterexamples. [https://people.eecs.berkeley.edu/~dawnsong/teaching/s10/pape...]
Mark E. Gold, Complexity of automaton identification from given data. [http://sebastian.doc.gold.ac.uk/papers/Language_Learning/gol...]
Mark E. Gold, Language Identification in the Limit.[http://web.mit.edu/6.863/www/spring2010/readings/gold67limit...]
Pitt, L., & Pitt, L. (1989). Inductive Inference, DFAs and Computational Complexity. 2ND INT. WORKSHOP ON ANALOGICAL AND INDUCTIVE INFERENCE (AII, 18--44. Retrieved from http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.187....
Leonard Pitt & Manfred K. Warmuth. The minimum consistent DFA problem cannot be approximated within any polynomial. [http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.86....]
Also see below for a gentle introduction to PAC-Learning, which actually allows for learning several concepts in restricted (but non-trivial) settings:
http://web.cs.iastate.edu/~honavar/pac.pdf
There is a lot more bibliography on both Inductive Inference tractability and PAC-Learnability. You'll have to follow the trail of references :)
Yes, it's entirely possible that we have missed something really important and when we figure that thing out we'll end up with a completely different Theory of Computation than we have now, and that may allow us to progress in learning concept classes that, today, are not learnable. That however, would be a revolution on par with Einstein's in physics.
Another thing that might save our collective ass, in terms of learnability, is that many of those results are assuming that P != NP. If that turns to be wrong, then those results will be off.
However, all this is very unlikely and in any case a) we can hardly expect such a profound re-interpretation of all that we know to happen within the next couple of generations and b) even if it does it remains the case that at this point in time everything we know suggests that there are some concept classes that we can simply not learn efficiently, at least not with computers.
Edit: what I'm really trying to say is that we may not understand human minds, but we're talking about computers here, so basically I agree with what you say, but with an added flourish :)
See deong's comment below. We may not understand how our minds work, but we do understand how computers work and that's where the learnability results come in: because they pertain to computers, not human minds.
Also, it may be that we don't need to learn every single thing we know. See Chomsky's Universal Grammar, etc.
I've been looking for work in this area without much success since publication is dominated by the opposite movement. I'd be grateful for any references you can dig up.
Love your point about hawking. I treat that as pure entertainment but since the general population treats hawking with high regard (well deserved) his opinions about things he doesn't know much about also carry weight in shaping public perception.
As far as speculating on how scary an AI driven future can be, a filmmaker in the 60s or a high school student today can do as compelling a job as Hawking.
I'm pretty sure Siri/Alexa/GoogleNow don't use rules. Like all modern Q&A systems like IBM Watson, they use a variety of statistical word match/associations and phrase/sentence response templates. Like Watson, the algorithm identifies the most probable word match to fill the void in your query and then inserts it into the most appropriate answer template. Yes, it's a bit hit or miss (like Watson answering "Toronto" for the US city with an airport named after a WWII hero). But if Apple/Google/Amazon adds some simple sequential disambiguation and refinement (clarification Q&A), I suspect this approach should answer most casual questions quite nicely.
IMHO, calling it "AI" means that the problem is not amenable to a closed form solution, so some form of iterative search and refinement is required. In venerable GOFAI parlance, AI problems are inherently 'scruffy'. Once the problem can be solved efficiently, it becomes 'neat' and it's no longer considered to be AI.
IMHO, most of today's progress in AI is scruffy (probabilistic word matches and deep learning), since they're driven not by formal internally-curated information but by smart search against presumed-relevant external referents, requiring secondary refinement -- a process that is largely inefficient and open-ended and error-prone, and will remain so indefinitely, I think, since we're not trying to improve the quality of the referent data, which is the weakest link in the cognitive chain.
Having said that, I still see abuse of the phrase (more in industry than academia) where standard machine learning applied to a workflow is marketed as some big AI in a product. But I understand why they do it - it usually helps in the long drawn out enterprise sales process to be perceived as an AI powered product.
Something like this would be immensely valuable, and would immediately improve healthcare everywhere, since the software could be updated with new or obscure knowledge right away, so you wouldn't have to shop around for doctors until you came across someone who had read of a very rare disease somewhere, the software would already know everything you might have.
Is there something like that out there? It sounds like there must be, but I don't know of any doctors using anything like this.
There are a lot of barriers to presenting it as you imagine, though: some liability, some workflow, and some just the deep lack of appropriate datasets. There is also some adoption issues - clinicians are rarely going to be happy with a black box approach but it can be difficult to explain/expose the details of such analysis in a way that clinicians will accept.
A last thought: this is actually very hard to do well and dangerous to make mistakes with - so your rather cavalier claim that there would be immediate and wide spread improvement of health care should be tempered a bit.
Another thing to consider in any work like this, you really need to thank about what your false positives mean, in a practical sense. If your numbers are big enough (e.g. screening) and your positives result in a procedure (e.g. biopsy), eventually someone is going to die based on a false positive. This may still be the better option, of course, but you really need to demonstrate that, and think about what your sensitivity and specificity tuning needs.
Going back further, medical diagnosis was a classic application of expert systems even in the 1970s. They ended up being something of a failure then due to the need to laboriously extract huge amounts of knowledge from doctors to get them to work, but it does show that it's long been a popular idea in the field.
MYCIN was never actually used in practice but research indicated that it proposed an acceptable therapy in about 69% of cases, which was better than the performance of infectious disease experts who were judged using the same criteria. [...] Some observers raised ethical and legal issues related to the use of computers in medicine — if a program gives the wrong diagnosis or recommends the wrong therapy, who should be held responsible? However, the greatest problem, and the reason that MYCIN was not used in routine practice, was the state of technologies for system integration, especially at the time it was developed. MYCIN was a stand-alone system that required a user to enter all relevant information about a patient by typing in responses to questions MYCIN posed. The program ran on a large time-shared system, available over the early Internet (ARPANet), before personal computers were developed. In the modern era, such a system would be integrated with medical record systems, would extract answers to questions from patient databases, and would be much less dependent on physician entry of information. In the 1970s, a session with MYCIN could easily consume 30 minutes or more — an unrealistic time commitment for a busy clinician.
Given that the systems integration problems should be soluble now, if not already solved, it does seem a little surprising that there isn't a modern version of such a thing, far more sophisticated, in general use. I guess the ethics and liability problems still loom larger than the above quote suggests.
[0] https://en.wikipedia.org/wiki/Mycin
[1] http://people.dbmi.columbia.edu/~ehs7001/Buchanan-Shortliffe...
Also such system is quite complex in many dimensions, so the natural risk minimization strategy would be to introduce it in stages: a system that warns about drug interactions here, monitors drug dosage there, reminds you to order extra tests, etc .And of course everything needs to be clinically tested to be accepted by docs(even though it isn't stricly necessary, it's just a second opinion, reminders, not the final word) . It makes for a very slow process.
And they they all face very big adoption barriers , in some cases upto 96% of their recommendations are skipped. Some of this is due to usability , but a lot due social/political factors.
Now after seeing this, what company in it's right mind will invest in developing a full diagnostical system ?
Actually , i know of one who did, http://www.isabelhealthcare.com , but at least according to the founding story , it began because the founder nearly lost his daughter to a hospital misdiagnosis, so he wasn't really driven by money.
And as far as i know(haven't followed them in a long time), they only achieved small scale sucsess.
The other company now entering this is IBM , but that's a bit of a unique case: IBM may be in big trouble so they take a huge bet, and the field they choose - cancer , is such that patients have much more power than usual , a patient has time, and can demand a diagnosis by Watson , heck it's even a cloud service, you just need to send blood and pay.
What's more, its definitely open to the public, except it isn't really promoted at all (it's only open in the interests of openness, you're not really meant to use it.) As such I can't find the URL for the life of me. Perhaps they closed it.
Any doctor or nurse will know about it, perhaps someone can post the link.
Lots, and for decades. If you Google "AI medical diagnosis" you get a zillion hits, including a dedicated Wikipedia page.
What we have right now is a boom in AI engineering, with AI research fairly neglected because it doesn't produce the same kind of hype.
Here's a Twitter thread [1] where NLP researchers describe people who work on machine translation as 'depressed', because they spend their time tweaking neural net parameters to get tiny improvements over the best techniques of last decade, with no opportunity to understand what causes an improvement.
It's well known in AI that common sense AI systems are incredibly hard to build and expert AI systems are relatively straight forward to model and build. Given that the most 'common sense' type products we see like Siri and Alexa still appear to be rules based systems (that have been enhanced with historical data no doubt), we all know the world is not suddenly going to become a common sense AI paradise unless a new Marvin Minsky makes some breakthroughs.
I find that the term 'AI' is frequently used when a company has taken an existing problem and enhanced the solution by using machine learning at one or more places. Nothing wrong with that fundamentally, but calling that AI is more of a marketing gimmick than anything else.
Google/Facebook do some of the best machine learning enhanced workflows that delight users. They usually refrain from branding it as AI.