The Computer Scientist Training AI to Think with Analogies
quantamagazine.org
quantamagazine.org
When the player's AI client completes a level it ends or "dies," it then "wakes up" into it's next level with a mostly blank slate and some basic models, and the object is to use the clues from what it percieves in its environment and the words you as a player give it, to find the next level. In the moment it finally apprehends its substrate again, it in effect "dies" from the perspective of the other client AIs in the game, and you get to play another round at the next level. There isn't really a way to end it because it's just fun, and more like an instrument you pick up and play than a finite game. The idea is the players aren't allowed to write clues directly into the game, but some of them will manage to cheat, and because it can go on forever, it's not like it will suddenly pop into reality, so the only point of it is the joy and fun of doing it.
Anyway, in this context, analogies would be like like mnemonics that loosely encode the isomorphic paths in a knowledge graph, which could be useful in the game, if being at risk of their ambiguity creating exploitable vulnerabilities for inserting direct clues, if one pursued it.
About a decade ago I cold-emailed them asking to do a postdoc with them to specifically fuse these concepts but it didn't really work out.
There is another theory which states that this has already happened.”
― Douglas Adams, The Restaurant at the End of the Universe
You can totally difference embedding vectors to represent relations. The canonical example from word2vec is basically solving an analogy. The big problem that you run into when applying this stuff more broadly is context, particularly how much context relating the subject and object you want, and which features encode that context. So the problem of abstraction she is talking about maps onto a regularization / feature selection problem.
More amusing is a comparison with a 5yo human child: https://twitter.com/lacker/status/1294341796831477761
A pity that she didn't mention any of that, it's pretty interesting.
> Q: If axbxcx goes to abc, what does xpxqxr go to?
> A: s
I've gotta be missing something here?
It should be noted that predictive coding predates Hawkins' work, perhaps best exemplified by Rao and Ballard's hierarchical predictive coding model from 1999. Various researchers working from similar assumptions put their thoughts together in Bayesian Brain in 2006. Since then, predictive processing has grown to be a very popular framework in neuroscience and cognitive science in general. Friston's free energy principle and philosopher Andy Clark's popularization of it have a lot to do with that.
There's been a ton of progress so far, so there's plenty of academic material for you to dive into.
Anyway, Hofstadter's (and his student Cabrera's) aim actually hasn't been to deliver a convincing industrial scale system but produce models that illuminate our understanding of thought. I'm not actually sure if that is the right way of doing it but it's worth noting they're not trying to "deliver" on the terms akin to today's neural networks.
[1] Understanding black box predictions through influence functions, https://arxiv.org/abs/1703.04730 [2] Evaluation of Similarity-based Explanations, https://arxiv.org/abs/2006.04528 [3] https://www.willows.ai/blog/getting-more
Do you have any good references besides Mitchell's Copycat and French's Tabletop books/thesis?
- He looks and sounds so much like Gilligan from Gilligan’s Island that I found it distracting.
- His ideas about analogies in computing were incredibly vague.
I am a big fan of the guy. Godel, Escher, Bach was the first book that really taught me what logic is. But it’s fascinating to reflect on how little computing has advanced with regard to practical analogizing.
GPT-3 is a parlor trick. Making vapid systems that fool people who aren’t thinking deeply is not going to end in successful general AI.
I predict, when the day comes that real general AI lands, we will discover that it is chronically manic or depressed or otherwise uncooperative. We will have re-invented teenagers. And we will never, ever be able to safely trust them.
I just saw a product being described as containing the “distilled wisdom of 1000 testers.” So, really? Are you actually wondering how people like me can think that some might be trying to work on genral AI?
AGI though is just the worst. If I am consider to posses general intelligence then why I am so bad at painting? I would love to be able to paint, even to just do Picasso ripoffs but I can't even make something close after much time, effort and training on past data/paintings.
When a computer has the same type of problem though then it negates any concept of general intelligence for the computer.
Nothing short of an Artificial God will satisfy what we mean by AGI.
Our poorly defined language here is causing philosophical problems.
I think a good analogy is to artificial light. The attitude towards AI would be to not be impressed by the light bulb but be constantly waiting for some miniature Sun/star that doesn't even make sense. That light bulb is not real artificial light says the AGI fanatic.
Some of my prior comments also talk about how this is same approach in martial arts. You are training techniques to develop muscle memory so you can apply them when a particular pattern appears. Then, a term I've used in the dojo and the classroom is that our goal is to "neutralize the attack/problem" into one of our pre-known templates.
[1] Soloway, E., & Ehrlich, K. (1984). Empirical studies of programming knowledge. - https://ieeexplore.ieee.org/abstract/document/5010283
I think most of our analogies are highly visual in their nature - we take the metaphorical and give it physical form, and through that mapping gain a common understanding of abstract topics. Our minds largely evolved around being able to make sense of and operate within the visual world, so it would make sense for our cognition to be tied into that ancient and massively powerful compute element. It’s embedded throughout human culture and language, so is likely not an emergent phenomenon but an inherent property of our minds.
On that basis, I wonder if we’re nearer than we think. I’m absolutely not an AI researcher, and it probably shows, but perhaps a visual intermediary is the key to generating a useful understanding of analogies in a way that maps well to the human understanding of analogies, and thus brings us a step closer to a general purpose AI.
Joshua Tenenbaum : https://mitibmwatsonailab.mit.edu/people/joshua-tenenbaum/
John Laird : https://laird.engin.umich.edu/
Steven Muggleton : http://wp.doc.ic.ac.uk/shm/