Preetum Nakkiran
@preetumnakkiran.bsky.social
ML Research @ Apple.
Understanding deep learning (generalization, calibration, diffusion, etc).
preetum.nakkiran.org
Understanding deep learning (generalization, calibration, diffusion, etc).
preetum.nakkiran.org
Another intern opening on our team, for a project I’ll be involved in (deadline soon!)
My colleague Shuangfei Zhai is looking for a summer research intern to work on improving TarFlow at Apple. If interested, send your CV to szhai at apple.com by this week.
February 25, 2025 at 2:05 AM
Another intern opening on our team, for a project I’ll be involved in (deadline soon!)
Reposted by Preetum Nakkiran
Last month I co-taught a class on diffusion models at MIT during the IAP term: www.practical-diffusion.org
In the lectures, we first introduced diffusion models from a practitioner's perspective, showing how to build a simple but powerful implementation from the ground up (L1).
(1/4)
In the lectures, we first introduced diffusion models from a practitioner's perspective, showing how to build a simple but powerful implementation from the ground up (L1).
(1/4)
February 20, 2025 at 7:19 PM
Last month I co-taught a class on diffusion models at MIT during the IAP term: www.practical-diffusion.org
In the lectures, we first introduced diffusion models from a practitioner's perspective, showing how to build a simple but powerful implementation from the ground up (L1).
(1/4)
In the lectures, we first introduced diffusion models from a practitioner's perspective, showing how to build a simple but powerful implementation from the ground up (L1).
(1/4)
Paper🧵 (cross-posted at X): When does composition of diffusion models "work"? Intuitively, the reason dog+hat works and dog+horse doesn’t has something to do with independence between the concepts being composed. The tricky part is to formalize exactly what this means. 1/
February 11, 2025 at 5:59 AM
Paper🧵 (cross-posted at X): When does composition of diffusion models "work"? Intuitively, the reason dog+hat works and dog+horse doesn’t has something to do with independence between the concepts being composed. The tricky part is to formalize exactly what this means. 1/
finally managed to sneak my dog into a paper: arxiv.org/abs/2502.04549
February 10, 2025 at 5:03 AM
finally managed to sneak my dog into a paper: arxiv.org/abs/2502.04549
nice idea actually lol: “Periodic cooking of eggs” : www.nature.com/articles/s44...
February 9, 2025 at 4:54 AM
nice idea actually lol: “Periodic cooking of eggs” : www.nature.com/articles/s44...
Reposted by Preetum Nakkiran
Reminder of a great dictum in research, one of 3 drilled into us by my PhD supervisor: "Don't believe anything obtained only one way", for which the actionable dictum is "immediately do a 2nd independent test of something that looks interesting before in any way betting on it". Its a great activity!
Grateful for my topic modeling and word embeddings training, which made me suspicious of any output that "looks good" but for which I haven't seen any alternative outputs that might also "look good."
Running a prompt and getting output that looks good isn't sufficient evidence for a paper.
Running a prompt and getting output that looks good isn't sufficient evidence for a paper.
February 5, 2025 at 8:17 PM
Reminder of a great dictum in research, one of 3 drilled into us by my PhD supervisor: "Don't believe anything obtained only one way", for which the actionable dictum is "immediately do a 2nd independent test of something that looks interesting before in any way betting on it". Its a great activity!
I’ve been in major denial about how powerful LLMs are, mainly bc I know of no good reason for it to be true. I imagine this was how deep learning felt to theorists the first time around 😬
February 4, 2025 at 5:25 PM
I’ve been in major denial about how powerful LLMs are, mainly bc I know of no good reason for it to be true. I imagine this was how deep learning felt to theorists the first time around 😬
Reposted by Preetum Nakkiran
Last year, we funded 250 authors and other contributors to attend #ICLR2024 in Vienna as part of this program. If you or your organization want to directly support contributors this year, please get in touch! Hope to see you in Singapore at #ICLR2025!
Financial Assistance applications are now open! If you face financial barriers to attending ICLR 2025, we encourage you to apply. The program offers prepay and reimbursement options. Applications are due March 2nd with decisions announced March 9th. iclr.cc/Conferences/...
ICLR 2024 Financial Assistance
iclr.cc
January 21, 2025 at 3:52 PM
Reposted by Preetum Nakkiran
The thing about "AI progress is hitting a wall" is that AI progress (like most scientific research) is a maze, and the way you solve a maze is by constantly hitting walls and changing directions.
January 18, 2025 at 3:45 PM
The thing about "AI progress is hitting a wall" is that AI progress (like most scientific research) is a maze, and the way you solve a maze is by constantly hitting walls and changing directions.
Reposted by Preetum Nakkiran
Thrilled to share the latest work from our team at
@Apple
where we achieve interpretable and fine-grained control of LLMs and Diffusion models via Activation Transport 🔥
📄 arxiv.org/abs/2410.23054
🛠️ github.com/apple/ml-act
0/9 🧵
@Apple
where we achieve interpretable and fine-grained control of LLMs and Diffusion models via Activation Transport 🔥
📄 arxiv.org/abs/2410.23054
🛠️ github.com/apple/ml-act
0/9 🧵
December 10, 2024 at 1:09 PM
Thrilled to share the latest work from our team at
@Apple
where we achieve interpretable and fine-grained control of LLMs and Diffusion models via Activation Transport 🔥
📄 arxiv.org/abs/2410.23054
🛠️ github.com/apple/ml-act
0/9 🧵
@Apple
where we achieve interpretable and fine-grained control of LLMs and Diffusion models via Activation Transport 🔥
📄 arxiv.org/abs/2410.23054
🛠️ github.com/apple/ml-act
0/9 🧵
Reposted by Preetum Nakkiran
📢 My team at Meta (including Yaron Lipman and Ricky Chen) is hiring a postdoctoral researcher to help us build the next generation of flow, transport, and diffusion models! Please apply here and message me:
www.metacareers.com/jobs/1459691...
www.metacareers.com/jobs/1459691...
Postdoctoral Researcher, Fundamental AI Research (PhD)
Meta's mission is to build the future of human connection and the technology that makes it possible.
www.metacareers.com
January 6, 2025 at 5:37 PM
📢 My team at Meta (including Yaron Lipman and Ricky Chen) is hiring a postdoctoral researcher to help us build the next generation of flow, transport, and diffusion models! Please apply here and message me:
www.metacareers.com/jobs/1459691...
www.metacareers.com/jobs/1459691...
Giving a short talk at JMM soon, which might finally be the push I needed to learn Lean…
January 3, 2025 at 6:30 PM
Giving a short talk at JMM soon, which might finally be the push I needed to learn Lean…
Just read this, neat paper! I really enjoyed Figure 3 illustrating the basic idea: Suppose you train a diffusion model where the denoiser is restricted to be "local" (each pixel i only depends on its 3x3 neighborhood N(i)). The optimal local denoiser for pixel i is E[ x_0[i] | x_t[ N(i) ] ]...cont
January 1, 2025 at 2:46 AM
Just read this, neat paper! I really enjoyed Figure 3 illustrating the basic idea: Suppose you train a diffusion model where the denoiser is restricted to be "local" (each pixel i only depends on its 3x3 neighborhood N(i)). The optimal local denoiser for pixel i is E[ x_0[i] | x_t[ N(i) ] ]...cont
Reposted by Preetum Nakkiran
LLMs dont have motives, goals or intents, and so they wont lie or deceive in order to obtain them. but they are fantastic at replicating human culture, and there, goals, intents and deceit abound. so yes, we should also care about such "behaviors" (outputs) in deployed systems.
December 26, 2024 at 7:05 PM
LLMs dont have motives, goals or intents, and so they wont lie or deceive in order to obtain them. but they are fantastic at replicating human culture, and there, goals, intents and deceit abound. so yes, we should also care about such "behaviors" (outputs) in deployed systems.
Reposted by Preetum Nakkiran
One #postdoc position is still available at the National University of Singapore (NUS) to work on sampling, high-dimensional data-assimilation, and diffusion/flow models. Applications are open until the end of January. Details:
alexxthiery.github.io/jobs/2024_di...
alexxthiery.github.io/jobs/2024_di...
December 15, 2024 at 2:46 PM
One #postdoc position is still available at the National University of Singapore (NUS) to work on sampling, high-dimensional data-assimilation, and diffusion/flow models. Applications are open until the end of January. Details:
alexxthiery.github.io/jobs/2024_di...
alexxthiery.github.io/jobs/2024_di...
Reposted by Preetum Nakkiran
“Should you still get a PhD given o3” feels like a weird category error. Yes, obviously you should still have fun and learn things in a world with capable AI. What else are you going to do, sit around on your hands?
December 22, 2024 at 4:29 AM
“Should you still get a PhD given o3” feels like a weird category error. Yes, obviously you should still have fun and learn things in a world with capable AI. What else are you going to do, sit around on your hands?
Catch our talk about CFG at the M3L workshop Saturday morning @ Neurips! I’ll also be at the morning poster session, happy to chat
December 14, 2024 at 6:56 AM
Catch our talk about CFG at the M3L workshop Saturday morning @ Neurips! I’ll also be at the morning poster session, happy to chat
Reposted by Preetum Nakkiran
Found slides by Ankur Moitra (presented at a TCS For All event) on "How to do theoretical research." Full of great advice!
My favourite: "Find the easiest problem you can't solve. The more embarrassing, the better!"
Slides: drive.google.com/file/d/15VaT...
TCS For all: sigact.org/tcsforall/
My favourite: "Find the easiest problem you can't solve. The more embarrassing, the better!"
Slides: drive.google.com/file/d/15VaT...
TCS For all: sigact.org/tcsforall/
December 13, 2024 at 8:31 PM
Found slides by Ankur Moitra (presented at a TCS For All event) on "How to do theoretical research." Full of great advice!
My favourite: "Find the easiest problem you can't solve. The more embarrassing, the better!"
Slides: drive.google.com/file/d/15VaT...
TCS For all: sigact.org/tcsforall/
My favourite: "Find the easiest problem you can't solve. The more embarrassing, the better!"
Slides: drive.google.com/file/d/15VaT...
TCS For all: sigact.org/tcsforall/
Reposted by Preetum Nakkiran
When a bunch of diffusers sit down and talk shop, their flow cannot be matched😎
It's time for the #NeurIPS2024 diffusion circle!
🕒Join us at 3PM on Friday December 13. We'll meet near this thing, and venture out from there and find a good spot to sit. Tell your friends!
It's time for the #NeurIPS2024 diffusion circle!
🕒Join us at 3PM on Friday December 13. We'll meet near this thing, and venture out from there and find a good spot to sit. Tell your friends!
December 12, 2024 at 1:15 AM
When a bunch of diffusers sit down and talk shop, their flow cannot be matched😎
It's time for the #NeurIPS2024 diffusion circle!
🕒Join us at 3PM on Friday December 13. We'll meet near this thing, and venture out from there and find a good spot to sit. Tell your friends!
It's time for the #NeurIPS2024 diffusion circle!
🕒Join us at 3PM on Friday December 13. We'll meet near this thing, and venture out from there and find a good spot to sit. Tell your friends!
I used to not really care to take pictures of myself, because what would I do with them, but over the break I had a fun training a bunch of Flux LoRAs on myself (surprisingly easy). And now I have renewed interest in profile photos.
December 6, 2024 at 12:46 AM
I used to not really care to take pictures of myself, because what would I do with them, but over the break I had a fun training a bunch of Flux LoRAs on myself (surprisingly easy). And now I have renewed interest in profile photos.
Reposted by Preetum Nakkiran
Help, I'm going to NeurIPS next week!
If you are presenting something that you think I should check out, pls drop the deets in thread (1-2 sentence description + where to find you)! 🙏
If you are presenting something that you think I should check out, pls drop the deets in thread (1-2 sentence description + where to find you)! 🙏
December 3, 2024 at 7:29 PM
Help, I'm going to NeurIPS next week!
If you are presenting something that you think I should check out, pls drop the deets in thread (1-2 sentence description + where to find you)! 🙏
If you are presenting something that you think I should check out, pls drop the deets in thread (1-2 sentence description + where to find you)! 🙏
Reposted by Preetum Nakkiran
An aspect of flow matching which I find a bit interesting is that it is covariant under affine changes of coordinate (c.f. optimal transport, which need not be). This allows for a few nice WLOGs, which I imagine have more applications than I realise.
Optimal transport computes an interpolation between two distributions using an optimal coupling. Flow matching, on the other hand, uses a simpler “independent” coupling, which is the product of the marginals.
December 2, 2024 at 1:21 PM
An aspect of flow matching which I find a bit interesting is that it is covariant under affine changes of coordinate (c.f. optimal transport, which need not be). This allows for a few nice WLOGs, which I imagine have more applications than I realise.
Reposted by Preetum Nakkiran
Blog post link: diffusionflow.github.io/
Despite seeming similar, there is some confusion in the community about the exact connection between the two frameworks. We aim to clear up the confusion by showing how to convert one framework to another, for both training and sampling.
Despite seeming similar, there is some confusion in the community about the exact connection between the two frameworks. We aim to clear up the confusion by showing how to convert one framework to another, for both training and sampling.
Diffusion Meets Flow Matching
Flow matching and diffusion models are two popular frameworks in generative modeling. Despite seeming similar, there is some confusion in the community about their exact connection. In this post, we a...
diffusionflow.github.io
December 2, 2024 at 6:45 PM
Blog post link: diffusionflow.github.io/
Despite seeming similar, there is some confusion in the community about the exact connection between the two frameworks. We aim to clear up the confusion by showing how to convert one framework to another, for both training and sampling.
Despite seeming similar, there is some confusion in the community about the exact connection between the two frameworks. We aim to clear up the confusion by showing how to convert one framework to another, for both training and sampling.
Sander poses a good question in this thread: Why is there so little variance in learning of diffusion models? Ie, why do different models trained on [random samples from] the same data distribution end up learning *nearly-identical* Noise --> Data function? Some background & speculations:
No, at this point I'm just using "diffusion models" and "flow matching" interchangeably 😁
I'm talking specifically about the observation that these models seem to recover the same (or similar) mappings every time, regardless of initialisation etc.
I'm talking specifically about the observation that these models seem to recover the same (or similar) mappings every time, regardless of initialisation etc.
November 30, 2024 at 6:55 PM
Sander poses a good question in this thread: Why is there so little variance in learning of diffusion models? Ie, why do different models trained on [random samples from] the same data distribution end up learning *nearly-identical* Noise --> Data function? Some background & speculations:
Reposted by Preetum Nakkiran
An intro to TOC class should indeed focus on "useful" ideas. But "useful" is often incredibly narrowlt defined. TOC is more way than its theorems, it's a way of modeling & thinking about problems that is applicable much more broadly across sciences & engg. Wigderson argues better than I could.
My take—certainly how I approach both PL and formal methods—is that we should focus on "the other 90%". There's 10% of the class who are just like us; you can teach them almost *anything*, and they'll get the rest in grad school. But those methods don't work, and even turn off, the 90%. ↵
November 28, 2024 at 1:04 PM
An intro to TOC class should indeed focus on "useful" ideas. But "useful" is often incredibly narrowlt defined. TOC is more way than its theorems, it's a way of modeling & thinking about problems that is applicable much more broadly across sciences & engg. Wigderson argues better than I could.