Slashdot Mirror
'Modern AI is Good at a Few Things But Bad at Everything Else' (wired.com)
Jason Pontin, writing for Wired: Sundar Pichai, the chief executive of Google, has said that AI "is more profound than ... electricity or fire." Andrew Ng, who founded Google Brain and now invests in AI startups, wrote that "If a typical person can do a mental task with less than one second of thought, we can probably automate it using AI either now or in the near future." Their enthusiasm is pardonable.
[...] But there are many things that people can do quickly that smart machines cannot. Natural language is beyond deep learning; new situations baffle artificial intelligences, like cows brought up short at a cattle grid. None of these shortcomings is likely to be solved soon. Once you've seen you've seen it, you can't un-see it: deep learning, now the dominant technique in artificial intelligence, will not lead to an AI that abstractly reasons and generalizes about the world. By itself, it is unlikely to automate ordinary human activities.
To see why modern AI is good at a few things but bad at everything else, it helps to understand how deep learning works. Deep learning is math: a statistical method where computers learn to classify patterns using neural networks. [...] Deep learning's advances are the product of pattern recognition: neural networks memorize classes of things and more-or-less reliably know when they encounter them again. But almost all the interesting problems in cognition aren't classification problems at all.
[...] But there are many things that people can do quickly that smart machines cannot. Natural language is beyond deep learning; new situations baffle artificial intelligences, like cows brought up short at a cattle grid. None of these shortcomings is likely to be solved soon. Once you've seen you've seen it, you can't un-see it: deep learning, now the dominant technique in artificial intelligence, will not lead to an AI that abstractly reasons and generalizes about the world. By itself, it is unlikely to automate ordinary human activities.
To see why modern AI is good at a few things but bad at everything else, it helps to understand how deep learning works. Deep learning is math: a statistical method where computers learn to classify patterns using neural networks. [...] Deep learning's advances are the product of pattern recognition: neural networks memorize classes of things and more-or-less reliably know when they encounter them again. But almost all the interesting problems in cognition aren't classification problems at all.
AI getting into the trough (https://en.wikipedia.org/wiki/Hype_cycle) again (https://en.wikipedia.org/wiki/AI_winter)?
Prominent people seem to fear AI (http://time.com/3614349/artificial-intelligence-singularity-stephen-hawking-elon-musk/), but isn't this just Fear of the Unknown? I mean, Elon and Stephen are really smart people, but do they know that most NN:s come down to linear algegra and spiced with non-linearities in the end, just simulating neurons? I mean neurons are common-place on the planet already, equipped with malice and stuff...
This. It makes sense that google will tout its neural networks, they own them. And yes, the reality is that many tasks and displays of "intelligence" will be difficult of those specific algorithms to handle efficiently or correctly. But the field is in its infancy. Computers haven't been around for even a century. I think though that they have in very specific terms been intelligent all along. The fact that they can do math such as understand 2+2=4 is in and of itself AMAZING.
Why it doesn't impress us is because we know what's going on inside and can dispell the magic. We know how it works. If I showed you a machine and I said "it can treat you like a therapist and cure your depression with greater success rate than the worlds renound phychiatrists", or some other seemingly "beyond computers" task; you would say that's artificial intelligence. But once I show you the secret sauce, the algorithm, the data points, the learning attributes it takes in and the process it uses, it's no longer intelligent, it's just a dumb machine using someone it was given. That's because we don't know why we are intelligent. We can use natural language, and we can do facial recognition, and we can determine creatively how to fix something we haven't seen before. We don't understand the process we take as toddlers to gain those skills. If we did, we would replicate it simply.
True AI will never become a reality because we have to understand it to build it, and by understanding it, we remove the magic and dispell that which was created as "true AI". We just keep moving the goal posts in search of something that is seemingly human. We will get there though. There is nothing in our heads that the universe and all of physics has barred us from creating. There is no law like gravity that states lIntelligence shall not exist but for within the head of a human being". Computers are better than us at chess, go, poker, and so many other tasks. Surely that is intelligence already.
If it's not better than a Human with an IQ of no less than 135 at literally everything it's not AI.
Why? We recognize and can measure intelligence in animals, so there is a wide range of non-human, natural intelligence that has been identified. Why would artificial intelligence have to start above all that?
Computers don't even have anything near mouse-level intelligence, we literally scanned (at the closest attempt) 1/100th of a mouse brain and simulated that, for a few milliseconds worth of time utilizing a massive supercomputer. We could likely use the existing algorithms to create AI, even per my definition of Human-level intelligence, but the hardware for it doesn't exist. If we dedicated every supercomputer on Earth to the task we couldn't even manage realtime simulation of a mouse brain - that's how pathetically lacking we are in the realm of AI. I'd place my bests on a vat-grown brain with electrodes attached before I would anything in silico.
...before a bunch of angry old coots post telling us that none of this is AI.
Let's put some of that into context.
A 5-year old can recognize a dog in an image in about 1/2 a second. A neuron takes about 0.05 seconds to activate and fire, so on average the entire recognition process takes about 10 steps.
Those steps include reading the image (sensing and converting the image data to internal form), and activating the physical response: saying "dog" or clicking the right button or whatever.
So let me ask this: what AI algorithm takes ten *steps* to recognize something as complicated as a dog?
Note that this works with dogs partially obscured (half masked by a tree, for instance), any size, rotated, from any angle (top down, face on, from the side), any breed (dalmations and chihuahuas), toys made to look like dogs, and cartoon dogs.
The algorithm does this at a very high level of accuracy, and can tell dogs apart from other animals with similar features: cats, opossums, and so on.
And the algorithm does this without a zillion training examples. A typical 5-year old has seen far fewer dogs than the Tensor Flow algorithm training set.
So tell me again: in what measure is our current level of AI anywhere close to being "real" AI?