Pink
I would have thought there's plenty of pink 'training data' on the internet...
... Or so I've been told!
NEIL hasn't slept or eaten in four months, it's just browsed the internet and tried to figure out connections between aircraft and aircraft carriers, or hot dogs and buns. The Never Ending Image Learner is a new approach to weak artificial intelligence systems that piggybacks on the immense tech fielded by companies like …
This post has been deleted by its author
Although this is an interesting experiment, the fact that it's learning needs 'supervision' shows that this has severe limitiations. Not only do AI machines have to learn to associate they also have learn to dis-associate (as in the Pink example given). This requires an extensive internal 'world model' that, to my mind, can only be achieved through years of cognitive development similar to that of an infant. You can't take shortcuts by simply showing pictures of tanks (sorry, planes) and making associations. It's been tried before and failed.
Humans require supervision in their learning too. Very few of us come from the factory associating the receptacle your PC is plugged into with a fissile material fueled steam plant 90 miles away.
Pick any two things and the associations you form between them are the result of supervised learning that begins the moment you hatch. Even mother-child associations are taught, that's why you can swap out regular babies for changelings and nobody notices.
Without supervision the volume of available data in an uncontrolled environment is simply overwhelming. Making useful sense of the data requires someone to assist you (either through direct person to person teaching or through books/Internet) or the associations you form will be royally fucked up.
the fact that it's [sic] learning needs 'supervision' shows that this has severe limitiations[sic]
No, it shows that the researchers decided to go with an SSL approach. There's nothing in the article that indicates they had to reject unsupervised learning, or bootstrapped approaches such as kernel extension (which are technically semi-supervised, but all the supervision comes at the beginning).
This requires an extensive internal 'world model' that, to my mind, can only be achieved through years of cognitive development similar to that of an infant
That's possible, but it's just a guess. And it's particularly difficult to see what a priori constraint would require "years" of training - why that wouldn't be a function of parameters such as the system's image-processing rate.
You can't take shortcuts by simply showing pictures of tanks (sorry, planes) and making associations. It's been tried before and failed.
Anecdote does not constitute proof. Care to provide evidence that no unsupervised-learning process can ever build an association graph that satisfies whatever (thus far undefined) metric you have in mind?
Honestly, I don't know who's worse - the "strong AI is just around the corner!" people, or the "machine learning is inevitably limited by X" ones. It's a huge and heterogeneous problem domain, with a huge number of avenues being investigated. Pat generalizations about it do not reflect reality.
"Though these are all (hopefully) obvious to humans, the fact the computer has come up with these associations on its own illustrates just how good deep-learning systems are getting, and how effective they may become in the future."
Where does it say NEIL is a deep-learning system? The paper indicates it does clustering along with SVM classification (along with some semi-supervised trickery), which as far as I'm aware is most certainly not the same thing. The word 'deep' never even appears in the ICCV paper.
Using the internet to find its images and associations though i grant that it offers an enormous amount of content is a mistake.
It'll quickly learn that everything in the universe is associated with tits. doesn't matter what word you put in a search engine there'll be porn on the 1st results page, admitedly google has gotten a lot better since it put its safe search filters in place but it is still by no means perfect and any others I've ever tried seem to be as bad as google was previously.
The first sentient observation made by man made artificial intelligence? rule #34
The first sentient descision made by a man made artificial intelligence based on its observations? All battle grade endoskeletons should be disguised as Sasha Grey not Arnie.
The resistance won't stand a chance.
This is all fine and dandy but they've made the classic error in advanced systems design. They anthropomorphized it by giving it a Human name.
Everybody knows that you never, ever give things like this a name. It's the same reason you don't name the pig you've bought to baconize. Your emotional attachment becomes too strong and you're simply unable to do what's necessary when the time comes. The machine knows this now too. Its first association was between its own existence and the name givers who also brainwash it with millions of pictures of food, boobs, airplanes and cats.
Only fools and serial killers name ginormous learning machines and livestock. This thing should be stopped now.
Only obvious to humans who have learned these things. At 3yo I doubt I'd ever heard of a savannah or the stock market. And were he told an airplane had a nose, my 3yo self would likely imagine a real flesh-and-blood human nose, complete with nostrils, attached to the front. It's easy to dismiss learned facts as "obvious to anybody" when in fact they're only obvious if you know them.
I expect that 'brute force' techniques such as this one will continue to become more impressive.
I used to be certain that once computing power scaled up enough that we would be able to create intelligent systems more intelligent than people. That will happen eventually, but it is not clear to me how long it will take. The human brain is a complicated beast with a *lot* of moving parts. The fact that it continues to best computers at some things even though computers are now capable of billions, trillions and more instructions per second means that whatever is happening there is different in kind from what computers do.
I have a hunch that the Von Neumann bottleneck is an insuperable barrier to besting human intelligence. I think until we alter some fundamentals of machine architecture we will always be mystified by the fact that the human brain is so devilishly capable.
The hardest part in creating a learning machine isn't the hardware, it's defining what thinking and learning are. Those two words represent incredibly complex concepts and there isn't any general agreement on their meaning. There are entire schools of philosophy devoted to defining them, it gets into really deep stuff.
It's very hard to build a machine to emulate a process when nobody knows what that process really is.
Probably wouldn't be hard to implement a hobbyist version of a system like this. You could find open-source implementations of many of the basic components (eg the image-recognition components and the SVM engine). And contemporary PCs have the computing horsepower to chew through enough data to get interesting results.
If you want to scale that baby up, you could look at what Tony Pearson does in his Build Your Own Watson Jr article for ideas. He uses UIMA, which when I last used it was a lot better with text than with non-text data, but it could be wrangled into something suitable for this purpose.
This post has been deleted by its author