Zuckerberg, Musk Invest in Artificial-Intelligence Company Vicarious

> I was referring to AIXI as the perfect mathematical AI.

I will have to Google AIXI. A big point about being mathematical is that that is about the only solid way we can evaluate candidate work before running software and, say, something like a Turing test.

Some math is most of why we know, well before any software is written, that (1) heap sort will run in n ln(n), (2) AVL trees find leaves in ln(n), and (3) our calculations for navigating a space craft to the outer planets will work. More generally, the math is 'deductive' in a severe and powerful sense and, thus, about the only tool we have to know well in advance of, say, writing a lot of software.

But math does not have 'truth' and, instead, needs hypotheses. So, for hypotheses, for some design for some software for AI, we need some. Enough hypotheses are going to be a bit tough to find. And math gives only some mathematical conclusions, and we will need to know that these are sufficient for AI; for that we will want, likely need, a sufficiently clear definition of AI, that is, something better than just an empirical test such as a Turing test or doing well on and IQ test. Tough challenge.

Instead of such usage of math, about all we have in AI for a 'methodology' is, (1) here I have some intuitive ideas I like, (2) with a lot of money I can write the code and, maybe, get it to run, and (3) trust me, that program can read a plane geometry book with all the proofs cut out and, then, fill in all the proofs or some such. So, steps (1) and (2) are, in the opinion of anyone else, say, DARPA, 'long shots', and (3) will be heavily in the eye of the beholder. The challenges of (1), (2), and (3) already make AI an unpromising direction.

> The main recent advancement in machine learning is deep learning. It's advanced the state of the art in machine vision and speech recognition quite a bit.

AI has been talking about 'deep knowledge' for a long time. That was, say, in a program that could diagnose car problems, 'knowledge' that the engine connected to the transmission connected to the drive shaft connected to the differential connected to the rear wheels or some such and, then, be able to use this 'knowledge' in 'reasoning' to diagnose problems. E.g., a vibration could be caused by worn U-joints. When I was in AI, when I worked in the field, there were plenty of people who saw the importance of such 'deep knowledge' but had next to nothing on really how to make it real.

For 'deep learning', the last I heard, that was tweaking the parameters 'deep' in some big 'neural network', basically a case of nonlinear curve fitting. Somehow I just don't accept that such a 'neural network' is nearly all that makes a human brain work; that is, I'd expect to see some promising 'organization' at a higher level than just the little elements for the nonlinear curve fitting.

E.g., for speech recognition, I believe an important part of how humans do it is to take what they heard, which is often quite noisy and by itself just not nearly enough, and compare it with what they know about the subject under discussion and, then, based on that 'background knowledge', correct the noisy parts of what they heard. E.g., if the subject is a cake recipe for a party for six people, then it's not "a cup of salt" but maybe a cup or two or three of flour. If the subject is the history of US presidents and war, then "I'll be j..." may be LBJ and "sson" maybe "Nixon". Here the speech recognition is heavily from a base of 'subject understanding'. An issue will be, how the heck does the human brain sometimes make such 'corrections' so darned fast.

For image recognition, the situation has to be in part similar but more so: I doubt that we have even a shot at image recognition without a 'prior library' of 'object possibilities': That is, if we are looking at an image, say, from satellite, of some woods and looking for a Russian tank hidden there, then we need to know what a Russian tank looks like so that we can guess what a hidden Russian tank would look like on the image so that we can, then, look for that on the image. Here we have to understand lighting, shadows, what a Russian tank looks like from various directions, etc. So, we are using some real 'human knowledge' of the real thing, the tank, we are looking for.

E.g., my kitty cat has a food tray. He knows well the difference between that tray and everything else that might be around it -- jug of detergent, toaster, bottle of soda pop, decorative vase, a kitchen timer. Then I can move his food tray, and he doesn't get confused at all. Net, what he is doing with image recognition is not just simplistic and, instead, has within it a 'concept' of his food tray, a concept that he created. He's not stupid you know!

So, I begin to conclude that for speech and image recognition, e.g., handwriting recognition, we need a large 'base' of 'prior human knowledge' about the 'subject area', e.g., with 'concepts', etc., before we start. That is, we need close to 'full, real AI' just to, say, do well reading any handwriting. From such considerations, I believe we have a very long way to go.

Broadly one of my first cut guesses about how to proceed would be to roll back to something simpler in two respects. First, start with brains smaller, hopefully simpler, than those of humans. Maybe start with a worm and work up to a frog, bird, ..., in a few centuries, a kitty cat! Second, start with the baby animal and see how it learns once it starts to as an egg, once it's born, what it gets from its mother, etc. So, eventually work up to software that could start learning with just "Ma ma' and proceed from there. But can't just start with humans and "Ma ma" because a human just born likely already has somehow built in a lot that is crucial we just don't have a clue about. So, start with worms, frogs, birds, etc.

Another idea for how to proceed is to try for just simple 'cognition' with just text and image input and just text output. E.g., start with something that can diagram English sentences and move from there to some 'understanding', e.g., have made progress enough with 'meaning' that, e.g., know when two sentences with quite different words and grammar really mean essentially the same thing and when they don't mean the same thing report why not and be able to revise one of the sentences so that the two do mean the same thing. So, here we are essentially assuming that AI has to stand on some capabilities with language -- do kitty cats have an 'internal language'? Hmm ...! If kitty cats don't have such an 'internal language', then I am totally stuck!

Then with some text and image input, the thing should be able to cook up a good proof of the Pythagorean theorem.

I can believe that some software can diagram English sentences or come close to it, but that is just a tiny start on what I am suggesting. The real challenge, as I am guessing, is to have the software keep track of and manipulate 'meaning', whatever the heck that is.

And I would anticipate a 'bootstrap' approach: Postulate and program something for doing such things with meaning, 'teach' it, and then look at the 'connections' it has built internally, say, between words and meaning, and also observe that the thing appears to work well. So, it's a 'bootstrap' because it works without our having any very good prior idea just why; that is, we could not prove in advance that it could work.

So, for kitty cat knowledge, have it understand its environment in terms of 'concepts' (part of 'meaning') hard, soft, strong, weak, hot, cold, and, then, know when it can use its claws to hold on to a soft, strong, not too hot or too cold surface, push out of the way a hard, weak obstacle, etc.

Maybe some such research direction could be made to work. But I'm not holding my breath waiting.

> I was referring to AIXI as the perfect mathematical AI.

At

http://wiki.lesswrong.com/wiki/AIXI

I looked it up: His 'decision theory' is essentially just stochastic optimal control. I've seen elsewhere claims that stochastic optimal control is a universal solution to the best possible AI. Of course, need some probability distributions; in some cases in practice, have those.

That reference also has

> Solomonoff’s theory of universal induction formally solves the problem of sequence prediction for unknown prior distribution.

Hmm? Then the text says that this solution is not computable -- sound bad!

Such grand, maybe impossible, things are not nearly the only way to exploit mathematics to know more about what the heck we are doing in AI, etc.

Keep in mind evolution managed to make strong AI, us, through pretty much blind, random mutations, and inefficient selection.

The thing about deep learning is that it's not just nonlinear curve fitting. It learns increasingly high level features and representations of the input. Recurrent neural networks have the power of a Turing machine. And stuff like dropout are really efficient at generalization. My favorite example is word2vec. Creating a representation for every English word. Subtracting "man" from "king" and adding "woman" gives the representation for "queen".

Speech recognition is moving that way. It outputs a probability distribution of possible words, and a good language model can use that to figure out what is most likely. But even a raw deep learning net should eventually learn those relationships. Same with image recognition. I think you'd be surprised at what is currently possible.

The future looks bright.

There is a degree of hype. They are really good at pattern recognition, maybe even superhuman on some problems and with enough training and data. But certainly they can't "think" in a normal sense or are a magical solution to the AI problem. And like everything in AI, once you understand how it actually works, it may not seem as impressive as it did at first.

>instead of somehow just 'storing' the data on the tank separately looks like rewiring the Intel processor when download a new PDF file instead of just putting the PDF file in storage.

Good analogy, but how would you even do that? One picture of a tank isn't enough to generalize. Is a tank any image colored green? Is it any object painted camouflage? Is it any vehicle that has a tube protruding from it?

In order to learn, you need a lot of examples, and you need to test a lot of different hypotheses about what a tank is. That's a really difficult problem.