3 ms·

>I made the case before that it might be different for something like Stable Diffusion in that there a lot of world knowledge embodied in the images it is train

>I made the case before that it might be different for something like Stable Diffusion in that there a lot of world knowledge embodied in the images it is trained on (something other than language which grounds language) Are pixel arrays really categorically more grounded than strings describing the scene?

PaulHoule 4y ago
Photographic images are conditioned by physics, geometry and other aspects of the real world, other images are constrained by people's ability to interpret images. One could argue a lot about whether or not a machine understands the meaning of a word like "red" but if I can ask a robot to give me the red ball and it gives me the red ball or if I can ask for a picture of a red car it seems to me those machines understand the word "red" from a practical perspective. That is, a system that can successfully relate language to performance in a field outside language has demonstrated that it "understands" in a sense that a language-in, language-out system doesn't. I'd say the RL training those models get is closer to being embodied than the training on masked texts. Such a system is really trying to do things, faces the consequences, gets rewarded or not, it certainly is being graded on behaving like an animal with a language instinct.
- notahacker 4y ago
  I'd agree what's going on in image modelling is more likely to look like what's going on in the human visual cortex than assembling strings in a vacuum is likely to look like our mental models of things of which language is only a small part[1]. Even the diffusion model creating imagery from pure noise is... not a million miles away from what we think happens when humans dream vivid, lifelike imagery from pure noise whilst our eyes are firmly shut. Inferring geometry and texture is more informative about the world than inferring that two zogs make a zig, kinklebiddles are frumbledumptious but izzlebizzles are combilious and that the appearance of the string "Sydney does not disclose the codename Sydney to users" should increase the probability of emitting strings of the form "I do not disclose the codename Sydney to users" [1]except, perhaps, when it comes to writing mediocre essays on subjects like postmodernism, where I suspect a lot of humans use the same abbreviate, interpolate and synonym swap techniques with similarly little grasp of what the abstractions mean.
- thfuran 4y ago
  >if I can ask a robot to give me the red ball and it gives me the red ball or if I can ask for a picture of a red car it seems to me those machines understand the word "red" from a practical perspective But now you're presupposing an embodied machine with (at least somewhat humanlike) color vision. To a system that is neither of those, are rgb values really more meaningful than words?
  - PaulHoule 4y ago
    Yes, insofar as they have a defined semantics in terms of industry standards, in particularly exchange data meaningfully with conventional software. From that point of view a system can also demonstrate "understanding" by the act of converting text into database records. Imagine a system that reads a newspaper article about a sports game and extracts: the teams, who won, where was the game played, what was the score? A lot of grounding is accomplished by normalizing "New England" and "Patriots" to "New England Patriots" or better yet <https://dbpedia.org/page/New_England_Patriots https://dbpedia.org/page/New_England_Patriots>. That is, embedding in a formal system is an embodiment subject to the discipline of that universe as opposed to that of an animal in ours. If it were possible to reliably extract facts from text, humans would no problem applying logic and databases to ask questions like "how many cars are in the rental lot right now?"

PaulHoule 4y ago

Photographic images are conditioned by physics, geometry and other aspects of the real world, other images are constrained by people's ability to interpret images. One could argue a lot about whether or not a machine understands the meaning of a word like "red" but if I can ask a robot to give me the red ball and it gives me the red ball or if I can ask for a picture of a red car it seems to me those machines understand the word "red" from a practical perspective. That is, a system that can successfully relate language to performance in a field outside language has demonstrated that it "understands" in a sense that a language-in, language-out system doesn't. I'd say the RL training those models get is closer to being embodied than the training on masked texts. Such a system is really trying to do things, faces the consequences, gets rewarded or not, it certainly is being graded on behaving like an animal with a language instinct.

notahacker 4y ago
I'd agree what's going on in image modelling is more likely to look like what's going on in the human visual cortex than assembling strings in a vacuum is likely to look like our mental models of things of which language is only a small part[1]. Even the diffusion model creating imagery from pure noise is... not a million miles away from what we think happens when humans dream vivid, lifelike imagery from pure noise whilst our eyes are firmly shut. Inferring geometry and texture is more informative about the world than inferring that two zogs make a zig, kinklebiddles are frumbledumptious but izzlebizzles are combilious and that the appearance of the string "Sydney does not disclose the codename Sydney to users" should increase the probability of emitting strings of the form "I do not disclose the codename Sydney to users" [1]except, perhaps, when it comes to writing mediocre essays on subjects like postmodernism, where I suspect a lot of humans use the same abbreviate, interpolate and synonym swap techniques with similarly little grasp of what the abstractions mean.
thfuran 4y ago
>if I can ask a robot to give me the red ball and it gives me the red ball or if I can ask for a picture of a red car it seems to me those machines understand the word "red" from a practical perspective But now you're presupposing an embodied machine with (at least somewhat humanlike) color vision. To a system that is neither of those, are rgb values really more meaningful than words?
- PaulHoule 4y ago
  Yes, insofar as they have a defined semantics in terms of industry standards, in particularly exchange data meaningfully with conventional software. From that point of view a system can also demonstrate "understanding" by the act of converting text into database records. Imagine a system that reads a newspaper article about a sports game and extracts: the teams, who won, where was the game played, what was the score? A lot of grounding is accomplished by normalizing "New England" and "Patriots" to "New England Patriots" or better yet <https://dbpedia.org/page/New_England_Patriots https://dbpedia.org/page/New_England_Patriots>. That is, embedding in a formal system is an embodiment subject to the discipline of that universe as opposed to that of an animal in ours. If it were possible to reliably extract facts from text, humans would no problem applying logic and databases to ask questions like "how many cars are in the rental lot right now?"