8 ms·

I'm only part way through the paper, but what struck me as interesting so far is this: In other text-to-image algorithms I'm familiar with (the ones you'll typ

by Imnimo 5y ago

I'm only part way through the paper, but what struck me as interesting so far is this: In other text-to-image algorithms I'm familiar with (the ones you'll typically see passed around as colab notebooks that people post outputs from on Twitter), the basic idea is to encode the text, and then try to make an image that maximally matches that text encoding. But this maximization often leads to artifacts - if you ask for an image of a sunset, you'll often get multiple suns, because that's even more sunset-like. There's a lot of tricks and hacks to regularize the process so that it's not so aggressive, but it's always an uphill battle. Here, they instead take the text embedding, use a trained model (what they call the 'prior') to predict the corresponding image embedding - this removes the dangerous maximization. Then, another trained model (the 'decoder') produces images from the predicted embedding. This feels like a much more sensible approach, but one that is only really possible with access to the giant CLIP dataset and computational resources that OpenAI has.

recuter 5y ago
What always bother me with this stuff is, well, you say one approach is more sensible than the other because the images happen to come out more pleasing. But there's no real rhyme or reason, it is a sort of alchemy. Is text encoding strictly worse or is it an artifact of the implementation? And if it is strictly worse, which is probably the case, why specifically? What is actually going on here? I can't argue that their results are not visually pleasing. But I'm not sure what one can really infer from all of this once the excitement washes over you. Blending photos together in a scene in photoshop is not a difficult task. It is nuanced and tedious but not hard, any pixel slinger will tell you. An app that accepts a smattering of photos and stitches them together nicely can be coded up any number of ways. This is a fantastic and time saving photoshop plugin. But what do we have really? "Kuala dunking basketball" needs to "understand" the separate items and select from the image library hoops and a Kuala where the angles and shadows roughly match. Very interesting, potentially useful. But if doesn't spit up exactly what you want can't edit it further. I think the next step has got to be that it conjures up a 3d scene in Unreal or blender so you can zoom in and around convincingly for further tweaks. Not a flat image.
- aaron695 5y ago
- ehsankia 5y ago
  > But there's no real rhyme or reason, it is a sort of alchemy. Is there a rhyme or reason as to why picasso decided to paint like that? Yes these networks are hard to reason about, but so are real human brains.
- woko 5y ago
  > Very interesting, potentially useful. But if doesn't spit up exactly what you want can't edit it further. DALL-E 2 spits as many outputs as you want. Then you choose the one you prefer.
- Imnimo 5y ago
  Yeah, I mean you're right that ultimately the proof is in the pudding. But I do think we could have guessed that this sort of approach would be better (at least at a high level - I'm not claiming I could have predicted all the technical details!). The previous approaches were sort of the best that people could do without access to the training data and resources - you had a pretrained CLIP encoder that could tell you how well a text caption and an image matched, and you had a pretrained image generator (GAN, diffusion model, whatever), and it was just a matter of trying to force the generator to output something that CLIP thought looked like the caption. You'd basically do gradient ascent to make the image look more and more and more like the text prompt (all the while trying to balance the need to still look like a realistic image). Just from an algorithm aesthetics perspective, it was very much a duct tape and chicken wire approach. The analogy I would give is if you gave a three-year-old some paints, and they made an image and showed it to you, and you had to say, "this looks like a little like a sunset" or "this looks a lot like a sunset". They would keep going back and adjusting their painting, and you'd keep giving feedback, and eventually you'd get something that looks like a sunset. But it'd be better, if you could manage it, to just teach the three-year-old how to paint, rather than have this brute force process. Obviously the real challenge here is "well how do you teach a three-year-old how to paint?" - and I think you're right that that question still has a lot of alchemy to it.
duxup 5y ago
This isn't something I'm knowledgeable on so forgive my simplification but is this like a sort of micro services for AI. Each AI takes their turn handing some aspect, another sort of mediates among them?
- Imnimo 5y ago
  I'd say Dall-E 2 is a little more unified - they do have multiple networks, but they're trained to work together. The previous approaches I was talking about are a lot more like the microservices analogy. Someone published a model (called CLIP) that can say "how much does this image look like a sunset". Someone else published a totally different model (e.g. VQGAN) that can generate images (but with no way to provide text prompts). A third person figures out a clever way to link the two up - have the VQGAN make an image, ask CLIP how much it looks like a sunset, and use backpropagation to adjust the image a little, repeat until you have a sunset. Each component is it's own thing, and VQGAN and CLIP don't know anything about one another.
  - astrange 5y ago
    VQGAN (being a "GAN") is already two networks - one Generates things, and the other is Adversarial and judges if the other network is good enough, then you train them both at once and they fight. CLIP+VQGAN generation IIRC works by replacing the adversarial network with CLIP, so it understands text prompts, then retraining it for a while towards the prompted target, then generating whatever it's learned from that. GANs are a silly idea that shouldn't work but somehow do. There's some attempts to replace the idea: https://www.microsoft.com/en-us/research/blog/unlocking-new-dimensions-in-image-generation-research-with-manifold-matching-via-metric-learning/ https://www.microsoft.com/en-us/research/blog/unlocking-new-...
    - Imnimo 5y ago
      I think that in CLIP+VQGAN, the VQGAN model is frozen, and what you do is start from a random latent code, generate an image, pass it to CLIP, and the backprop through CLIP and through the VQGAN generator to figure out how you should move the latent code to make it better match the prompt. Then you just keep taking gradient ascent steps to find better and better latent codes. So it's like 'retraining', except you're 'training' the network input rather than the network weights.
swalsh 5y ago
Do you think some of these techniques could be slightly modified, and applied to DNA sequences?
- j7ake 5y ago
  The problem is that with DNA sequences you dont have good training sets. With text and images you can leverage “ground truth” data (verified by humans) to train your model. The DNA sequences I would look for methods that don’t require good ground truth data.
- snek_case 5y ago
  Maybe very very short (single-gene) sequences. The thing with DNA is it's the product of evolution. The DNA guides the synthesis of proteins, then the proteins fold into a 3D shape, and they interact with chemicals in their environment based on their shape. In the context of a living being, different genes interact with each other as well. For example, you have certain cells that secrete hormones (many genes needed to do that), then you have genes that encode for hormone receptors, and those receptors trigger other actions encoded by other genes. There's probably too much complexity to ask an AI system to synthesize the entire genetic code for a living being. That would be kind of like if I asked you to draw the exact blueprints for a fighter get, and write all the code, and synthesize all the hardware all at once, and you only get one shot. You would likely fail to predict some of the interactions and the resulting system wouldn't work. You could only achieve this through an iterative process that would involve years of extensive testing. Could you use a deep learning system to synthesize genetic code? Maybe just single genes that do fairly basic things, and you would need a massive dataset. Hard to say what that would look like. Is it really enough to textually describe what a gene does?
  - Jack000 5y ago
    This is all true, but it doesn't preclude the possibility of generating DNA. Human share a lot of DNA sequences with other animals, and the genetic differences between individual humans are even smaller. You might have trouble generating a human with horns or something, but a taller one is probably mostly an engineering problem. What GPT-3 and DALL-E shows is that you can infer a lot based on the latent structure of data, even without understanding the underlying physical process.
krick 5y ago
While the whole narrative of your comment totally makes sense, I don't really see the difference between the two approaches, not on a conceptual level. You still needed to train this so called "prior" at some point (so, I'm also not sure if it's fair to call it a "prior"). I mean, the difference between your two descriptions seems to be the difference between descriptions (i.e., how you chose to name individual parts of the system), not the systems. I'm not sure if I'm speaking clearly, I just don't understand, what's the difference between training "text encoding to an image" vs "text embedding to image embedding". In both cases you have some kind of "sunset" (even though it's obviously just a dot in a multi-dimension space, not the letters) on the left, and you try to maximize it when training the model to get either a image-embedding or a image straight away.
- Imnimo 5y ago
  Yeah, my comment didn't really do a good job of making clear that distinction. Obviously the details are pretty technical, but maybe I can give a high-level explanation. The previous systems I was talking about work something like this: "Try to find me the image the looks like it most matches 'a picture of a sunset'. Do this by repeatedly updating your image to make it look more and more like a sunset." Well, what looks more like a sunset? Two sunsets! Three sunsets! But this is not normally the way images are produced - if you hire an artist to make you a picture of a bear, they don't endeavor to create the most "bear" image possible. Instead, what an artist might do is envision a bear in their head (this is loosely the job of the 'prior' - a name I agree is confusing), and then draw that particular bear image. But why is this any different? Who cares if the vector I'm trying to draw is a 'text encoding' or an 'image encoding'? Like you say, it's all just vectors. Take this answer with a big grain of salt, because this is just my personal intuitive understanding, but here's what I think: These encodings are produced by CLIP. CLIP has a text encoder and an image encoder. During training, you give it a text caption and a corresponding image, it encodes both, and tries to make the two encodings close. But there are many images which might accompany the caption "a picture of a bear". And conversely there are many captions which might accompany any given picture. So the text encoding of "a picture of a bear" isn't really a good target - it sort of represents an amalgamation of all the possible bear pictures. It's better to pick one bear picture (i.e. generate one image embedding that we think matches the text embedding), and then just to try to draw that. Doing it this way, we aren't just trying to find the maximum bear picture - which probably doesn't even look like a realistic natural image. Like I said, this is just my personal intuition, and may very well be a load of crap.

by Imnimo 5y ago

recuter 5y ago

What always bother me with this stuff is, well, you say one approach is more sensible than the other because the images happen to come out more pleasing. But there's no real rhyme or reason, it is a sort of alchemy. Is text encoding strictly worse or is it an artifact of the implementation? And if it is strictly worse, which is probably the case, why specifically? What is actually going on here? I can't argue that their results are not visually pleasing. But I'm not sure what one can really infer from all of this once the excitement washes over you. Blending photos together in a scene in photoshop is not a difficult task. It is nuanced and tedious but not hard, any pixel slinger will tell you. An app that accepts a smattering of photos and stitches them together nicely can be coded up any number of ways. This is a fantastic and time saving photoshop plugin. But what do we have really? "Kuala dunking basketball" needs to "understand" the separate items and select from the image library hoops and a Kuala where the angles and shadows roughly match. Very interesting, potentially useful. But if doesn't spit up exactly what you want can't edit it further. I think the next step has got to be that it conjures up a 3d scene in Unreal or blender so you can zoom in and around convincingly for further tweaks. Not a flat image.

aaron695 5y ago
ehsankia 5y ago
> But there's no real rhyme or reason, it is a sort of alchemy. Is there a rhyme or reason as to why picasso decided to paint like that? Yes these networks are hard to reason about, but so are real human brains.
woko 5y ago
> Very interesting, potentially useful. But if doesn't spit up exactly what you want can't edit it further. DALL-E 2 spits as many outputs as you want. Then you choose the one you prefer.
Imnimo 5y ago
Yeah, I mean you're right that ultimately the proof is in the pudding. But I do think we could have guessed that this sort of approach would be better (at least at a high level - I'm not claiming I could have predicted all the technical details!). The previous approaches were sort of the best that people could do without access to the training data and resources - you had a pretrained CLIP encoder that could tell you how well a text caption and an image matched, and you had a pretrained image generator (GAN, diffusion model, whatever), and it was just a matter of trying to force the generator to output something that CLIP thought looked like the caption. You'd basically do gradient ascent to make the image look more and more and more like the text prompt (all the while trying to balance the need to still look like a realistic image). Just from an algorithm aesthetics perspective, it was very much a duct tape and chicken wire approach. The analogy I would give is if you gave a three-year-old some paints, and they made an image and showed it to you, and you had to say, "this looks like a little like a sunset" or "this looks a lot like a sunset". They would keep going back and adjusting their painting, and you'd keep giving feedback, and eventually you'd get something that looks like a sunset. But it'd be better, if you could manage it, to just teach the three-year-old how to paint, rather than have this brute force process. Obviously the real challenge here is "well how do you teach a three-year-old how to paint?" - and I think you're right that that question still has a lot of alchemy to it.

duxup 5y ago

This isn't something I'm knowledgeable on so forgive my simplification but is this like a sort of micro services for AI. Each AI takes their turn handing some aspect, another sort of mediates among them?

Imnimo 5y ago
I'd say Dall-E 2 is a little more unified - they do have multiple networks, but they're trained to work together. The previous approaches I was talking about are a lot more like the microservices analogy. Someone published a model (called CLIP) that can say "how much does this image look like a sunset". Someone else published a totally different model (e.g. VQGAN) that can generate images (but with no way to provide text prompts). A third person figures out a clever way to link the two up - have the VQGAN make an image, ask CLIP how much it looks like a sunset, and use backpropagation to adjust the image a little, repeat until you have a sunset. Each component is it's own thing, and VQGAN and CLIP don't know anything about one another.
- astrange 5y ago
  VQGAN (being a "GAN") is already two networks - one Generates things, and the other is Adversarial and judges if the other network is good enough, then you train them both at once and they fight. CLIP+VQGAN generation IIRC works by replacing the adversarial network with CLIP, so it understands text prompts, then retraining it for a while towards the prompted target, then generating whatever it's learned from that. GANs are a silly idea that shouldn't work but somehow do. There's some attempts to replace the idea: https://www.microsoft.com/en-us/research/blog/unlocking-new-dimensions-in-image-generation-research-with-manifold-matching-via-metric-learning/ https://www.microsoft.com/en-us/research/blog/unlocking-new-...
  - Imnimo 5y ago
    I think that in CLIP+VQGAN, the VQGAN model is frozen, and what you do is start from a random latent code, generate an image, pass it to CLIP, and the backprop through CLIP and through the VQGAN generator to figure out how you should move the latent code to make it better match the prompt. Then you just keep taking gradient ascent steps to find better and better latent codes. So it's like 'retraining', except you're 'training' the network input rather than the network weights.

swalsh 5y ago

Do you think some of these techniques could be slightly modified, and applied to DNA sequences?

j7ake 5y ago
The problem is that with DNA sequences you dont have good training sets. With text and images you can leverage “ground truth” data (verified by humans) to train your model. The DNA sequences I would look for methods that don’t require good ground truth data.
snek_case 5y ago
Maybe very very short (single-gene) sequences. The thing with DNA is it's the product of evolution. The DNA guides the synthesis of proteins, then the proteins fold into a 3D shape, and they interact with chemicals in their environment based on their shape. In the context of a living being, different genes interact with each other as well. For example, you have certain cells that secrete hormones (many genes needed to do that), then you have genes that encode for hormone receptors, and those receptors trigger other actions encoded by other genes. There's probably too much complexity to ask an AI system to synthesize the entire genetic code for a living being. That would be kind of like if I asked you to draw the exact blueprints for a fighter get, and write all the code, and synthesize all the hardware all at once, and you only get one shot. You would likely fail to predict some of the interactions and the resulting system wouldn't work. You could only achieve this through an iterative process that would involve years of extensive testing. Could you use a deep learning system to synthesize genetic code? Maybe just single genes that do fairly basic things, and you would need a massive dataset. Hard to say what that would look like. Is it really enough to textually describe what a gene does?
- Jack000 5y ago
  This is all true, but it doesn't preclude the possibility of generating DNA. Human share a lot of DNA sequences with other animals, and the genetic differences between individual humans are even smaller. You might have trouble generating a human with horns or something, but a taller one is probably mostly an engineering problem. What GPT-3 and DALL-E shows is that you can infer a lot based on the latent structure of data, even without understanding the underlying physical process.

krick 5y ago

While the whole narrative of your comment totally makes sense, I don't really see the difference between the two approaches, not on a conceptual level. You still needed to train this so called "prior" at some point (so, I'm also not sure if it's fair to call it a "prior"). I mean, the difference between your two descriptions seems to be the difference between descriptions (i.e., how you chose to name individual parts of the system), not the systems. I'm not sure if I'm speaking clearly, I just don't understand, what's the difference between training "text encoding to an image" vs "text embedding to image embedding". In both cases you have some kind of "sunset" (even though it's obviously just a dot in a multi-dimension space, not the letters) on the left, and you try to maximize it when training the model to get either a image-embedding or a image straight away.

Imnimo 5y ago
Yeah, my comment didn't really do a good job of making clear that distinction. Obviously the details are pretty technical, but maybe I can give a high-level explanation. The previous systems I was talking about work something like this: "Try to find me the image the looks like it most matches 'a picture of a sunset'. Do this by repeatedly updating your image to make it look more and more like a sunset." Well, what looks more like a sunset? Two sunsets! Three sunsets! But this is not normally the way images are produced - if you hire an artist to make you a picture of a bear, they don't endeavor to create the most "bear" image possible. Instead, what an artist might do is envision a bear in their head (this is loosely the job of the 'prior' - a name I agree is confusing), and then draw that particular bear image. But why is this any different? Who cares if the vector I'm trying to draw is a 'text encoding' or an 'image encoding'? Like you say, it's all just vectors. Take this answer with a big grain of salt, because this is just my personal intuitive understanding, but here's what I think: These encodings are produced by CLIP. CLIP has a text encoder and an image encoder. During training, you give it a text caption and a corresponding image, it encodes both, and tries to make the two encodings close. But there are many images which might accompany the caption "a picture of a bear". And conversely there are many captions which might accompany any given picture. So the text encoding of "a picture of a bear" isn't really a good target - it sort of represents an amalgamation of all the possible bear pictures. It's better to pick one bear picture (i.e. generate one image embedding that we think matches the text embedding), and then just to try to draw that. Doing it this way, we aren't just trying to find the maximum bear picture - which probably doesn't even look like a realistic natural image. Like I said, this is just my personal intuition, and may very well be a load of crap.