4 ms·

Anyone who has kids and teaches them things will know AI learns and ‘understands’ very differently. I can say something like ‘a tiger is just a lion with strip

Anyone who has kids and teaches them things will know AI learns and ‘understands’ very differently. I can say something like ‘a tiger is just a lion with stripes’ to a 3 year old and they now ‘understand’ what a tiger is almost as well as if they saw a picture of one. They could definitely identify one from a picture now. This kind of understanding won’t work with an AI because we don’t understand what characteristics it has latched into when identifying a lion. For all we know it’s that the background of each lion image its been trained on has a blue sky. Or the tigers are all looking at the camera. I think the ability to pick apart what you know and learn new things by reasoning about that knowledge is the key to if understanding is taking place.

ShamelessC 5y ago
I've done a fair bit of work with multimodal deep learning and I am fairly confident that a DALL-E, CLIP, or NUWA architecture would output/classify those phrases accurately without being trained explicitly on images of Tigers. I see your point however.
- jonplackett 5y ago
  I’d be interested in where to look for more info on that. Do you think it could work on anything more complex? As I said in comment below I reckon I could make a much more elaborate explanation and still have the kid get it.
  - ShamelessC 5y ago
    https://openai.com/blog/dall-e/ https://openai.com/blog/dall-e/ shows a decent ability to generalize to previously unseen concepts. You're correct to consider the complexity of the phrase and just how good humans are at this sort of thing without needing much "training". For now, concepts that aren't explicitly in the training set are effectively composed from those which are. This can lead to some bizarre and outright incorrect results, particularly when it comes to counting objects in a scene or with relative positioning between objects (e.g. a blue box on top of a red rectangle to the left of a green triangle) but it's early days and there's lots of progress happening all the time.
    - jonplackett 5y ago
      Thanks. this is interesting. In a way it's opposite thought-direction of what we are talking about. eg. can it look at an avocado shaped chair and recognise it as a chair in the shape of an avocado - for me that would display a lot more understanding of the concept of 'chair' and 'avocado' than being able to produce an image of the phrase 'a chair shaped like an avocado' - but maybe the same process must be happening in there somewhere to make this possible? What do you think?
      - ShamelessC 5y ago
        https://openai.com/blog/clip/ https://openai.com/blog/clip/ CLIP is the corrolary model created purely for classification purposes rather than generative as in DALL-E and is quite impressive across a range of tasks. Give it an image and a caption, and in return you get a score (0.0 to 1.0) telling you how much they match. I think it is more in line with your premise. Others have taken CLIP and combined it with frozen language models (GPT2) to create automatic captioning models that are very impressive. edit: To try to address your question about whether or not actual semantic composition occurs, I think the answer is "yes" but it would be challenging to convince you this is true without going into details of the "self attention" mechanism which allows both methods to work. The short version is that these networks are able to find meaning in extremely high-dimensional problems by having a mechanism specifically tasked with learning positional statistics of the training data. In language this refers to e.g. how often the word "pillow" is directly next to the word "fort". In vision, this similarly refers to how often e.g. trees are positioned next to gift-wrapped presents. That's quite simplified but I hope that makes sense to a degree!
        jonplackett 5y ago
        Thanks for writing such detailed reply! Very helpful. It’s exciting that this is happening. Do these models still have to be trained in the same way and then become essentially static? What I also find very different about teaching a child is how dynamically their understanding can shift based on explanations and observations. That’s probably the another hallmark of understanding.
yboris 5y ago
I'm curious, when you say "This kind of understanding won’t work with an AI", do you mean currently, or in principle, even in the future? note: children's brains come pre-loaded with so much stuff when we are born (we are not "blank slates").
- jonplackett 5y ago
  What I mean is AIs aren’t really built with the goal of ‘understanding’ anything currently. They are awesome at individual tasks but they don’t have the kind of common sense a person can use to reason with and build up an understanding of how pieces of knowledge fit together. Eg. Driverless cars can identify a car, or a motorbike or a cyclist and maybe work out a trajectory for it. But they don’t understand that a bike is a person on top of a metal frame and that person is made up of a head and body and limbs. And if that head is looking away from them it can’t see them coming. For me, that’s understanding. Deconstructing and reconstructing knowledge to come to conclusions that add to your knowledge.
  - yboris 5y ago
    Thank you for the response. A bit of a push back: On the driverless cars example you're right that AI models abstract away needless complexity, but so do we, as humans, can live without understanding that bodies are made up of cells (or how muscles move during a tennis match). If we want AI models to include a person's gaze in their calculations, we can make it happen. From an abstract-enough perspective, the AI model will do enough that whether it "understands" may become irrelevant. Language understanding is much harder, because words are about stuff in the world. Just like when a toddler says "love" we know they don't fully understand what they mean, AI won't have the capacity to mean "love" unless it has a lot more it "understands" along the way. But it feels like it could in the near few decades "understand" enough about "duck" to mean it when it says "I see a duck".
    - jonplackett 5y ago
      Thanks for yours too! Isn’t this civilised? I can tell I’m not on reddit… I think what I’m getting at is the way an AI learns is kind of a black box. Maybe it is learning about cyclists head movements. Maybe it only learned that for cyclists wearing helmets. We don’t really know it’s put the pieces together correctly - like the classic example of training an AI to recognise friendly tanks and it just learning the difference between day and night. Because we let it essentially define its own characteristics for recognising things we don’t know if it picked the right ones. We just have to hope the training data was large enough to average out. Using the right features is fundamental to understanding what’s really going on. Perhaps occasionally an AI might notice some connection we humans have been ignoring and that might be good. But in general I think we’ve done a pretty good job of allocating words and language to the important things. We separated out the body into parts that make sense. That kind of thing. If an AI had those underlying building blocks wired up more like our own. maybe understanding in a similar way that we do would be a lot easier. Last quick example / thought experiment: To test understanding in a person you try asking them questions that would confuse them if they didn’t understand. Eg. If I photoshopped a crocodile head onto a Tiger’s body and have it a fish tail and put it on the moon. My toddle would still recognise all the bits and describe it as such. An AI would just come up 33% ish for each possibility because it doesn’t understand what an animal is and that it has those parts.
addsidewalk 5y ago
Pick apart is what they’re doing. The systems I’ve worked in immediately abstract strings, shapes in images, etc, into the mathematical shape and gaps between edges. If you dig into an arbitrary array in a variety of places, the fields contains coordinates, not “Hi Mom, kids are ok, blah blah”. It’s measuring the white space in a thing, where everything but the feature you’re currently interested in is white space; what’s between the features I want? Then comparing that to results of other data structures that had the same white space measuring applied. Does it not do what you said you do you not want to believe it? I think the issue is the companies being incredibly disingenuous about how this all works. At the root is elementary information theory: https://www.amazon.com/dp/0486240614 https://www.amazon.com/dp/0486240614 Formal language is 5,000 years old. Human intuition for quantitative assessment of hunger, warmth, supply stocks, tool building, etc is much older. IMO human language is noise obscuring obviousness. It’s the desktop metaphor of cognition. “Please internalize my language versus observe for yourself.”
bpizzi 5y ago
> I can say something like ‘a tiger is just a lion with stripes’ to a 3 year old and they now ‘understand’ what a tiger is almost as well as if they saw a picture of one. They could definitely identify one from a picture now. Assuming the 3 year old already knew what a lion looks like, and point at 'things with stripes' and 'things without stripes'. I think that a model that can already recognize separately lions and stripes should be able to tag a tiger's picture as a 'Lion with stripes', no?
- mannykannot 5y ago
  Not necessarily, if the stripes prevent tigers from scoring highly on the lion measure. Generalizing from what is formally insufficient information is something that humans are quite good at (though obviously not infallibly.)
- jonplackett 5y ago
  Maybe… but this is just one very easy example and also using something very obvious and visual. I could also say “a Cheetah is like a lion but it’s smaller and has spots and runs a lot faster. And a leopard is like a lion but smaller and can climb trees and has spots.” I could probably start with a house cat and describe an elephant if I wanted to and I’ll bet the kid would work it out. The ability to take apart and reassemble knowledge is what I’m talking about here, not just add two simple bits of information together.
  - justinpombrio 5y ago
    > I could also say “a Cheetah is like a lion but it’s smaller and has spots and runs a lot faster. And a leopard is like a lion but smaller and can climb trees and has spots.” The OpenAI website is unresponsive at the moment, so I can't actually demonstrate this, but you could totally tell GPT-3 that, and it would then make basic inferences. For example, saying "four" when asked how many legs a cheetah has, or guessing a smaller weight for a cheetah than a lion when asked to guess a specific weight for both. Not perfectly, but a lot better than chance, for the basic inferences. (You wouldn't actually tell it "a Cheetah is like a lion but..." because it already knows what a Cheetah is. Instead you'd say "a Whargib is like a lion but ...", and ask it basic questions about Whargibs.)

ShamelessC 5y ago

I've done a fair bit of work with multimodal deep learning and I am fairly confident that a DALL-E, CLIP, or NUWA architecture would output/classify those phrases accurately without being trained explicitly on images of Tigers. I see your point however.

jonplackett 5y ago
I’d be interested in where to look for more info on that. Do you think it could work on anything more complex? As I said in comment below I reckon I could make a much more elaborate explanation and still have the kid get it.
- ShamelessC 5y ago
  https://openai.com/blog/dall-e/ https://openai.com/blog/dall-e/ shows a decent ability to generalize to previously unseen concepts. You're correct to consider the complexity of the phrase and just how good humans are at this sort of thing without needing much "training". For now, concepts that aren't explicitly in the training set are effectively composed from those which are. This can lead to some bizarre and outright incorrect results, particularly when it comes to counting objects in a scene or with relative positioning between objects (e.g. a blue box on top of a red rectangle to the left of a green triangle) but it's early days and there's lots of progress happening all the time.
  - jonplackett 5y ago
    Thanks. this is interesting. In a way it's opposite thought-direction of what we are talking about. eg. can it look at an avocado shaped chair and recognise it as a chair in the shape of an avocado - for me that would display a lot more understanding of the concept of 'chair' and 'avocado' than being able to produce an image of the phrase 'a chair shaped like an avocado' - but maybe the same process must be happening in there somewhere to make this possible? What do you think?
    - ShamelessC 5y ago
      https://openai.com/blog/clip/ https://openai.com/blog/clip/ CLIP is the corrolary model created purely for classification purposes rather than generative as in DALL-E and is quite impressive across a range of tasks. Give it an image and a caption, and in return you get a score (0.0 to 1.0) telling you how much they match. I think it is more in line with your premise. Others have taken CLIP and combined it with frozen language models (GPT2) to create automatic captioning models that are very impressive. edit: To try to address your question about whether or not actual semantic composition occurs, I think the answer is "yes" but it would be challenging to convince you this is true without going into details of the "self attention" mechanism which allows both methods to work. The short version is that these networks are able to find meaning in extremely high-dimensional problems by having a mechanism specifically tasked with learning positional statistics of the training data. In language this refers to e.g. how often the word "pillow" is directly next to the word "fort". In vision, this similarly refers to how often e.g. trees are positioned next to gift-wrapped presents. That's quite simplified but I hope that makes sense to a degree!
      - jonplackett 5y ago
        Thanks for writing such detailed reply! Very helpful. It’s exciting that this is happening. Do these models still have to be trained in the same way and then become essentially static? What I also find very different about teaching a child is how dynamically their understanding can shift based on explanations and observations. That’s probably the another hallmark of understanding.

yboris 5y ago

I'm curious, when you say "This kind of understanding won’t work with an AI", do you mean currently, or in principle, even in the future? note: children's brains come pre-loaded with so much stuff when we are born (we are not "blank slates").

jonplackett 5y ago
What I mean is AIs aren’t really built with the goal of ‘understanding’ anything currently. They are awesome at individual tasks but they don’t have the kind of common sense a person can use to reason with and build up an understanding of how pieces of knowledge fit together. Eg. Driverless cars can identify a car, or a motorbike or a cyclist and maybe work out a trajectory for it. But they don’t understand that a bike is a person on top of a metal frame and that person is made up of a head and body and limbs. And if that head is looking away from them it can’t see them coming. For me, that’s understanding. Deconstructing and reconstructing knowledge to come to conclusions that add to your knowledge.
- yboris 5y ago
  Thank you for the response. A bit of a push back: On the driverless cars example you're right that AI models abstract away needless complexity, but so do we, as humans, can live without understanding that bodies are made up of cells (or how muscles move during a tennis match). If we want AI models to include a person's gaze in their calculations, we can make it happen. From an abstract-enough perspective, the AI model will do enough that whether it "understands" may become irrelevant. Language understanding is much harder, because words are about stuff in the world. Just like when a toddler says "love" we know they don't fully understand what they mean, AI won't have the capacity to mean "love" unless it has a lot more it "understands" along the way. But it feels like it could in the near few decades "understand" enough about "duck" to mean it when it says "I see a duck".
  - jonplackett 5y ago
    Thanks for yours too! Isn’t this civilised? I can tell I’m not on reddit… I think what I’m getting at is the way an AI learns is kind of a black box. Maybe it is learning about cyclists head movements. Maybe it only learned that for cyclists wearing helmets. We don’t really know it’s put the pieces together correctly - like the classic example of training an AI to recognise friendly tanks and it just learning the difference between day and night. Because we let it essentially define its own characteristics for recognising things we don’t know if it picked the right ones. We just have to hope the training data was large enough to average out. Using the right features is fundamental to understanding what’s really going on. Perhaps occasionally an AI might notice some connection we humans have been ignoring and that might be good. But in general I think we’ve done a pretty good job of allocating words and language to the important things. We separated out the body into parts that make sense. That kind of thing. If an AI had those underlying building blocks wired up more like our own. maybe understanding in a similar way that we do would be a lot easier. Last quick example / thought experiment: To test understanding in a person you try asking them questions that would confuse them if they didn’t understand. Eg. If I photoshopped a crocodile head onto a Tiger’s body and have it a fish tail and put it on the moon. My toddle would still recognise all the bits and describe it as such. An AI would just come up 33% ish for each possibility because it doesn’t understand what an animal is and that it has those parts.

addsidewalk 5y ago

Pick apart is what they’re doing. The systems I’ve worked in immediately abstract strings, shapes in images, etc, into the mathematical shape and gaps between edges. If you dig into an arbitrary array in a variety of places, the fields contains coordinates, not “Hi Mom, kids are ok, blah blah”. It’s measuring the white space in a thing, where everything but the feature you’re currently interested in is white space; what’s between the features I want? Then comparing that to results of other data structures that had the same white space measuring applied. Does it not do what you said you do you not want to believe it? I think the issue is the companies being incredibly disingenuous about how this all works. At the root is elementary information theory: https://www.amazon.com/dp/0486240614 https://www.amazon.com/dp/0486240614 Formal language is 5,000 years old. Human intuition for quantitative assessment of hunger, warmth, supply stocks, tool building, etc is much older. IMO human language is noise obscuring obviousness. It’s the desktop metaphor of cognition. “Please internalize my language versus observe for yourself.”

bpizzi 5y ago

> I can say something like ‘a tiger is just a lion with stripes’ to a 3 year old and they now ‘understand’ what a tiger is almost as well as if they saw a picture of one. They could definitely identify one from a picture now. Assuming the 3 year old already knew what a lion looks like, and point at 'things with stripes' and 'things without stripes'. I think that a model that can already recognize separately lions and stripes should be able to tag a tiger's picture as a 'Lion with stripes', no?

mannykannot 5y ago
Not necessarily, if the stripes prevent tigers from scoring highly on the lion measure. Generalizing from what is formally insufficient information is something that humans are quite good at (though obviously not infallibly.)
jonplackett 5y ago
Maybe… but this is just one very easy example and also using something very obvious and visual. I could also say “a Cheetah is like a lion but it’s smaller and has spots and runs a lot faster. And a leopard is like a lion but smaller and can climb trees and has spots.” I could probably start with a house cat and describe an elephant if I wanted to and I’ll bet the kid would work it out. The ability to take apart and reassemble knowledge is what I’m talking about here, not just add two simple bits of information together.
- justinpombrio 5y ago
  > I could also say “a Cheetah is like a lion but it’s smaller and has spots and runs a lot faster. And a leopard is like a lion but smaller and can climb trees and has spots.” The OpenAI website is unresponsive at the moment, so I can't actually demonstrate this, but you could totally tell GPT-3 that, and it would then make basic inferences. For example, saying "four" when asked how many legs a cheetah has, or guessing a smaller weight for a cheetah than a lion when asked to guess a specific weight for both. Not perfectly, but a lot better than chance, for the basic inferences. (You wouldn't actually tell it "a Cheetah is like a lion but..." because it already knows what a Cheetah is. Instead you'd say "a Whargib is like a lion but ...", and ask it basic questions about Whargibs.)