Comment by jbob2000 | Hacker News Reader

Comment by jbob2000 | Hacker News Reader

One thing Julian Jaynes discusses in his book Origin of Consciousness in the Breakdown of the Bicameral Mind (highly recommend, but a hard read) is that consciousness is comprised of a number of different qualities, one being excerptification - the ability to take a number of indiscriminate ideas/memories/images and string together a narrative.

For example, I can take the idea of a narwhal horn and the idea of a horse and mix them together to create a unicorn. My conscious mind took the excerpt of a narwhal horn and strung it together with the excerpt of a horse body. And I didn't just stick the horn on some random part of the horse's body, there was a contextual understanding behind it - a narwhal horn belongs on the forehead, the head of a horse is the thing with eyes and a nose.

Can a computer do that? No way, not presently at least. There's waaayy too much going on in the process of excerptification for a computer to replicate just this one facet of our consciousness.

J.C.R. Licklider, father of the Internet, thought analogy tests were a strong indication of intelligence. Hoeflin's difficult IQ tests also used analogies, but it actually now seems like a bad way to measure intelligence in a multicultural setting (for example "Jekyll : Hyde :: Eloi : ?" is a particularly pointless test of having "read" two particular mainstream books in English.)

Excerptification, as you describe it, sounds like a game of analogies. AI is not far from solving this kind of problem, as long as it is expressed symbolically.

Here's a paper discussing one AGI system, implemented in Haskell, that might be appropriate for work with analogy problems that require "imagination" as well as raw "perception":

Claes Strannegård; Rickard von Haugwitz; Johan Wessberg; Balkenius Christian (2013). "A Cognitive Architecture Based on Dual Process Theory" http://gup.ub.gu.se/publication/178583

Computers are close to that. The neural net image backward inference is similar. They take bits of imagery and try to retrofit it in 'meaningful' ways, blending both source and target image. Also genetic programming can model swapping and blending traits.

Meh, not really, the computer doesn't understand the images it's mixing, it's just using some clever algorithms to piece together an image. It has no awareness of the images you're giving it. I couldn't give it an image of a narwhal and an image of a horse and expect it to give me a unicorn.

To be fair, you haven't defined awareness in any meaningful way. Concepts are just coded by specific patterns of neural firings - there is no reason a neural net can't theoretically also code concepts meaningfully (of course, understanding that process and getting it work perfectly is decades away, IMO).

I don't buy the "understanding is hard". Also you'd have to characterize precisely what is so special about a unicorn. There are animals with 2, 3 .. things on their forehead. That doesn't make them special. What about pegasus. Or other invented beings of 50/50 blends (minautor, egyptian gods). We just have one particular model or animal qualities ingrained and 'enjoy' toying with a maximizing function of having both our traits and other animal traits. I'd love to see a cockroachman god figure.

I'd bet a thousands dollars that the underlying principles Deep Dreams uses to infer fractal and continuous images are the same we have in our brains.

> Meh, not really, the computer doesn't understand the images it's mixing,

That's 100% wrong. It does create abstract representations and it can imagine back. For example, there is this neural net that can do "woman + glasses = woman wearing glasses" on images, without any tags. It can combine concepts and generate new images. There is also another one that can "dream" images of bedrooms. It can even interpolate 10 intermediary bedrooms between any two images, and they look very good.

> I couldn't give it an image of a narwhal and an image of a horse and expect it to give me a unicorn.

You absolutely could with a generative adversarial network and the right training dataset. They can interpolate between images in a semantic way, and add and subtract components of the semantic representation. While it's not narwhals and horses, here's an example of using images of men with and without glasses along with a woman without glasses to generate a woman with glasses: https://raw.githubusercontent.com/Newmu/dcgan_code/master/im...

More results here: https://github.com/Newmu/dcgan_code And note that this field is advancing so rapidly that these results have already been improved on despite being only six months old.

jbob2000 makes the error of being too vague (didn't define awareness). You're making the error of being too specific. Under no suitable definition of awareness do GANs exhibit awareness. They are a very specific and narrow statistical tool used to solve a specific problem they are trained for specifically.

If you're referring to backprop, no. It's an open research question to characterize what/how a neural net is learning during backprop (aside from trivial statements like "optimizing a nonconvex function" etc.)

No, I didn't even study NN, so backprop is a fuzzy notion. I take it as a convergence feedback loop, am I completely off ?

By backward I meant the images done by the Deep Dreams program. They retrofit (hence the back) 'dog' signals fractally. This is a pretty amazing primitive to me. Trying to recognize 'concepts' even through that much distortion.

The proof is in the pudding. If it is useful.

People aren't useful. Hammers are useful. It seems hard to go from useful -> intelligence, and hard to go from intelligent -> conscious.

You mean like this? https://www.instapainting.com/ai-painter

Deep Dreams. Even though this one was surprisingly effective, but less general than DD.

Reply on news.ycombinator.com