I'm only about... 50% serious when I say this. But all of these huge tech companies getting into serious AI and learning really makes me worry that Singularity is coming way sooner then I would have hoped. Like in the next 10 years.
You could technically say it's a step above brute-forcing in that, after randomly assigning neural weights, gradient descent and other tricks are used to more quickly settle into good solutions. But the science is simply nowhere near the level of the human brain, and the current state of the art is still only "proficient" at relatively simple pattern matching, rather than coming up with decisions that actually influence the state of the world that it is perceiving.
The deep architectural insights that will likely be needed before computers can approach the level of human intelligence still haven't been made yet, it seems to me based on what I know of the current research.
While it's entirely likely that better AI algorithms will become available, the progress being made by "brute forcing" problems using lots of data and lots of computing power is much more than was expected 10 years ago.
For example, Facebook's facial recognition software now correctly recognizes faces 97.25% of the time. Humans do 97.53%[1].
I'd encourage you to read the Norvig et. al paper: "The Unreasonable Effectiveness of Data"
[1] http://www.technologyreview.com/news/525586/facebook-creates...
[2] https://static.googleusercontent.com/media/research.google.c...
To be clear, I'm not saying that new AI paradigms are needed to improve that facial recognition score from 97.25% to 99+% (the existing techniques might actually be optimal anyway). I'm saying they are needed to even be able to attempt other tasks that actual humans find effortless, such as interpreting literature, creating a song, carrying on a conversation, decorating a room in an appealing way, understanding symbolic math, having this debate with you, etc. (Certainly symbolic math software has been developed, but no artificial intelligence has been developed that "understands" it and can put it into larger context, particularly one based on neural networks a la the real human brain.)
I don't think my thinking is outdated (or even that controversial). It's mostly informed by a Deep Learning class I recently took taught by Yann LeCun, who is definitely at the forefront of the field (although these opinions are entirely my own).
EDIT: Also, when I mention new "AI paradigms", I do expect them to build on the deep neural networks currently being studied. But more sophisticated architectures than feedforward nets or even the simple recurrent nets that I've seen so far will be needed, I think. And that will likely require some breakthroughs in theory as well as engineering.
One could say that is only data, nothing "intelligent" about pattern-matching it.
But I'm biased, I'm a non-believer when it comes to reaching the singularity, as in I don't think we will ever be able to build something that will understand jokes, sexual innuendo or who/which will experience the Madeleine effect (http://en.wikipedia.org/wiki/Madeleine_%28cake%29#Literary_r...)
Before you use pattern recognition (i.e. pre-defined, remembered, or improvised abstraction for invariants) to solve a problem, you have to understand there's a problem to be solved.
ML seems to be doing well at pattern recognition, but that second problem - understanding what needs to be done in a situation, and why, and accurately predicting possible developments and consequences - is possibly more critical for real singularity-style intelligence.
It's machine reasoning as opposed to machine learning - in the street-wise sense, not the theorem proving sense.
Computers will certainly be able to understand jokes or innuendo. At least as well as people do, anyway.
As far as the Madeleine effect goes, I'm not sure why you consider that intelligence?
Serious question, are we close to anywhere on that? I last looked at AI-related stuff around 10 years ago, when "sentiment analysis" as it was then called was only capable of approximating what "good press" and "bad press" meant.
> As far as the Madeleine effect goes, I'm not sure why you consider that intelligence?
I'm making a generalization in thinking that "intelligence" ultimately means "human intelligence", and I'm also assuming (probably wrong, there's no exact science for it) that things like the "Madeleine effect" sit at the core of us being humans, and the third (and last generalization) in thinking that if you're not experiencing things like the "Madeleine effect" you're not really human-like, so you cannot really be "intelligent". Writers like Asimov and Arthur C. Clarke have written about this problem (what it means to be human when "robots" will come about) way better than I could ever dream of doing, but my favorite presentation of the potential issues when the Singularity will come around remains Stanislaw Lem's "The Cyberiad". Or we could continue focusing on technicalities only and call it a day.
While the book isn't about AI per se, the complexity that it hints at in biological systems is (yet another) warning shot across the bow of my personal AI voyaging ship, telling me to avoid those waters (actually it's more like the Kraken from Pirates of the Caribbean).
I've never espoused AI technology that attempted to emulate the physical brain, believing it too complex an approach. But taking another brief peek into the black box of living systems scared the crap out of me! I'm glad I didn't pursue biochemistry as a field of study so many years ago. My mind would be full up by now with little to show for it.
This is obviously not true AI, but powerful image recognition is one of the holy grails of the original AI quest. Reaching human level (and likely above-human level soon) image recognition capabilities will not unleash the singularity, but I don't think anyone knows how many steps lie between achieving this sub-step and achieving the final goal. Certainly visual recognition is a huge component of mammalian brains (both in terms of brain volume and in terms of behavior). Even more worrying, Microsoft claims that this same technology can be applied to sound recognition, and it doesn't seem unlikely that it could cover all senses, given proper inputs. If this all holds up and gets polished, this effectively marks the first major step in human-like AI. Again, how many steps remain no one knows, but we should absolutely take notice of results like this and seek better arguments one way or the other. Saying "nuh-uh" won't change the way this research unfolds, and does us no favors for preparing for it as it matures.
I understand that we are nowhere near building AI's that can model the human brain and thought process accurately. But I do think that there will be a point, far before that, at which point machines / ai's will have the ability to improve themselves outside of the algorithms programmed into them.
At that point, I think there is definitely the possibility for an exponentially growing increase in their sophistication, far far faster then humans have the ability to learn and develop. This is due to the fact that AI's never have to rest, never get tired or forget things, and could potentially have the entire world of networked devices and information at its disposal.
Anyway, I pretty much accepted that I was probably going to tank some down votes when I made my original post. But I achieved my main goal of seeing some interesting points / discussion about the feasibility of such a claim.