Re:The AI Misinformation Epidemic
by eli_gottlieb on Tuesday March 28, 2017 @01:45AM
(#13974405)
> It's still very unclear what, if anything, this algorithm has in common with how a human brain actually operates.
Not very much at all. Most psychological evidence shows that human beings seem to operate off composable, hierarchical generative models and probabilistic inference. Deep neural networks are basically just huge continuous circuit approximators.
>composable, hierarchical generative models and probabilistic inference. Deep neural networks are basically just huge continuous circuit approximators.
These two things are not mutually exclusive.
Re:The AI Misinformation Epidemic
by eli_gottlieb on Tuesday March 28, 2017 @01:33PM
(#13978670)
What sort of neural network implements human-style learning and inference?
numenta's HTM
While I don't take exception with the word machine learning (it's reasonably well defined), I agree with most of the points here. Some are grossly exaggerated (e.g. needing petabytes of data to do machine learning). In my research on medical time series, we get strong results with a few thousand examples (megabytes). Same for natural language datasets which are modestly sized. But yes, absolutely agree that function approximation for supervised learning and some kind of artificial consciousness as portrayed in the media are very far apart.
yep, call function-fitting/optimization "learning" and it's suddenly a form of intelligence!
What is human intelligence, if not the result of a colossal, distributed function optimisation process driven by evolutionary forces over millions of years?
true, it's the result of an optimization process, but intelligence itself (whatever it is...it's a suitcase-term) isn't as simple as that. Or maybe it is, I don't know!
"Optimization" describes the result, not the implementation. Unless you consider intelligence to be average-suboptimal, it's entirely reasonable to say that it's about optimization without committing you to any claim about how intelligence is implemented.
The unsupervised learning capabilities demonstrated by a human infant are a pretty lofty target. Give it time.
Lets say an infant is 3 years old, thats 26,280 hours, lets say a baby sleeps about 2/3 of the time - lets call it 8600 hours of awake time.
8600 hours of 4k video at 60fps (i would say this is lower quality then the human eye) is 2.657 petabytes.
Thats not including the sense of touch or smell, which i am sure are comparable in the amount of data processed by a human.
We're looking at 5+ petabytes of data that a human has to learn from by age 3 - Whats wrong with taking petabytes of data in order to learn meaningful things?
Unfortunately, the reality is that.... sorry... it's mostly just statistical algorithms based around regression and intermediate calculus. State-of-the-art "deep" neural networks are not really anything like the absurdly parallel, asynchronous biological networks that power our neo-cortex - rather, they're basically an application of matrix multiplication designed to "learn" a function by iteratively minimizing an error value using gradient descent. It's still very unclear what, if anything, this algorithm has in common with how a human brain actually operates. It turns out that these kind of statistical algorithms can work pretty well when you have petabytes of data to learn from. But we're still not anywhere near the unsupervised learning capabilities demonstrated by a human infant.