Video Of The Week: Why Toddlers Are Smarter Than Computers
I saw Gary Marcus give a talk at the NYU AI Event I blogged about this past week. In that
In that talk, he suggested that Deep Learning wasn’t going to get us all the way to where we want to get in AI.
I thought it was an interesting take on AI, particularly right now when the buzz and hype is so high.
This TedX talk makes the same argument and so I am sharing it with you.
Comments (Archived):
Testing.
one, two
Of course, the AI isn’t as smart as humans.Yesterday I was in a “Learning AI workshop” at Galvanize in SF. They’re the leading trainers of data scientists and AI folks for Bay Area startups.This is what I posted on the Slack Channel after the speaker said something about how babies can “just absorb a lot of information and cluster it without any supervision, labeling or guidance” and that’s how Deep Learning works.He has a PhD from Stanford but, clearly, can’t remember being a baby and his mother defining, contextualizing, labeling and risk managing concepts and language for him. https://uploads.disquscdn.c…Most male AI PhDs seem to be either unaware or in denial that the female X gene has been found by Cambridge University researchers to be a key source of intelligence wrt reasoning, language understanding and planning.The lack of female engineers (and ones with maths degrees who understand the deeper connections with Bayes, Gauss, Bernoulli, Euler et al, and who have other reference points to the sciences, Liberal Arts and languages) is how brogrammers end up not having their most basic wrong assumptions about artificial and human intelligence tested, by the way. https://uploads.disquscdn.c…So I sit in on “Learning AI” workshops in SF to ensure 175+ male engineers become aware how VITAL female intelligence is to their intelligence, and this is why they have to do much better and make AI that’s unbiased and more representative of that female and other diverse intelligences.
Much of sorting and grouping small data is learning the most abstract things and building from there. This talk is a great example of another proponent of toddlers as a needed group for more study to learn how we learn, for its own sake a key knowledge, but clearly at the service of deep learning and the augmented anythings.Working in this space, from the point of view of visual-spatial cognition and play is fascinating, it lives at the root of human experience.
This Saturday an AI instructor (with Stanford PhD) said, “Babies just do unsupervised learning. They don’t need any training.”Naturally, I posted this on the workshop’s slack channel to inform the 175 male engineers that their assumptions are completely wrong. https://uploads.disquscdn.c…The unfortunate thing about Silicon Valley and tech industry’s lack of inclusion and female engineers is that unscientific and non-sensical assumptions like “Babies can do clustering, pattern recognition and Deep Learning” get propagated.That and a handful of male PhDs build models according to their mental biases — rather than building systems that enable representation for billions of people.If anyone can argue that the definitions of a few dozen guys are a representative sample population of 3.7 billion people online, then the Pope isn’t Catholic.The various philosophical and psychology frameworks that these male PhDs have inherited may also hold them back from being able to develop a universal Natural Language Understanding model. https://uploads.disquscdn.c…@sigmaalgebra:disqus @wmoug:disqus — The problem’s much more complex than mere maths and we need multi-disciplinary problem-solvers to work on it.Otherwise, we’re in a situation where Alexa team sends me an email that says Echo can’t recognize my instruction “Alexa, tell me a da Vinci quote” and only recognizes “Alexa, ask da Vinci quotes to tell me a da Vinci quote.”For NLU, the AI has to understand free-form, non-deterministic phrases — not only specifically ordered commands that fit the code.
Toddlers smarter than computers?Sure.And, sure, what is apparently in current work in deep learning would seem to have essentially nothing to do with how natural intelligence works in kitty cats, puppy dogs, human babies, and human adults.Yes, if we want to program something like human intelligence, then we have to be masochistic not to notice how humans learn back to conception, at least back birth, and certainly for the first three years or so. We’re talking a lot of progress in just three years, about 1000 days. If focus on just the best 10 hours of each day, we’re talking just 10,000 hours.And THAT’S the big data input. With just that data, that’s how human babies do it.If that is too much data, then see how fast kitty cats, puppy dogs, horses, chimpanzees, octopi learn. Net — darned fast.IIRC at one point here I posted that every mother sees how her baby learns and sees such learning more clearly than we see in current AI discussions.On how to program real AI? I posted some first cut thoughts athttp://avc.com/2016/08/chec…http://avc.com/2016/10/sxsl…