Jiaming Song @baaadas Twitter profile | Pikagi

Pikagi

Jiaming Song

@baaadas

6,240

Followers

1,085

Following

88

Media

688

Statuses

Chief Scientist @LumaLabsAI . #LumaDreamMachine

https://t.co/tTjDZNELdU

Joined November 2014

Don't wanna be here? Send us removal request.

Pinned Tweet

@baaadas

Jiaming Song

4 months

Extremely proud to be working on this with many amazing people @LumaLabsAI ! Generate a 5-second, 120 frames video in 120 seconds from text or images *now* on: Available to everyone. #LumaDreamMachine

Tweet card media

Luma Dream Machine | AI Video Generator

Unlock your creativity with Luma AI Video Generator. Turn text into stunning videos with our cutting-edge text-to-video AI. Dream big, create bigger!

@LumaLabsAI

Luma AI

4 months

Introducing Dream Machine - a next generation video model for creating high quality, realistic shots from text instructions and images using AI. It’s available to everyone today! Try for free here #LumaDreamMachine

592

2K

7K

17

17

190

Last Seen Profiles

@lingons_boys

@ShiftyVincenzo

@keremtural6

@Dr_Alpha_sloth

@lacinut

@Dakoholics

@kuri0404

@isusemasaviral

@rehg_j

@markvdboogaard

@fashion_newbie_

@DNARZA7

@kr_maynard

@unxpjs

@jocelynnakiaa

@stwmaniax

@octophrator

@KontrolGrip

@DeVPrice_

@FatmaAnwer44511

@krbchnk

@DalilaFCoficial

@stwmaniax

@SewaneeXC_TF

@Sayngelic

@motimotidaihuke

@JFedpost1776

@peakbuns

@BinorRaja

@Grupo_CMG

@mamaniji_

@harkuladebiri

@Necksplitter

@fr33_tim3

@mediavida5

@lovesunwing1

@baaadas

Jiaming Song

3 years

"The paper is not novel because some arxiv paper in February of 2022 already did it" -- I recall that the #icml2022 submission deadline was on January of 2022? I am fine with the paper getting rejected, but not all of us have time machines 🤣 @icmlconf

Tweet media one

19

26

543

@baaadas

Jiaming Song

2 years

📢 We are looking for highly motivated #ML #AI Ph.D. students to work with us at NVIDIA Research as summer #interns next year. We encourage applicants with experience in generative modeling in one of these domains:

7

41

231

@baaadas

Jiaming Song

1 year

@chenlin_meng , @ArashVahdat , and I are presenting the #diffusion model tutorial at #CVPR2023 on June 18 (). Since there are > 1300 papers on this topic, we cannot read all of them😭, and we need your help on uncovering all the "hidden gems"!

Tweet card media

Denoising Diffusion-based Generative Modeling: Foundations and Applications

Tutorial on Diffusion Models in Conjunction with CVPR 2023

cvpr2023-tutorial-diffusion-models.github.io

8

51

223

@baaadas

Jiaming Song

1 year

📢 Our #CVPR2023 tutorial on "Denoising Diffusion Models: A Generative Learning Big Bang" w/ @chenlin_meng and @ArashVahdat is happening tomorrow morning! 9:00 to 12:30, West 202-204. This is the year of big bang for diffusion models in CVPR!

Tweet media one

6

37

197

@baaadas

Jiaming Song

1 year

After a wonderful year at NVIDIA, I am starting a new adventure @LumaLabsAI 🐻

16

8

204

@baaadas

Jiaming Song

3 years

Introducing Diffusion-Denoising Models with Contrastive Representations (D2C), a non-adversarial image generative model for few-shot conditional generation (e.g. image manipulation). w/ @a7b2_3 @chenlin_meng @StefanoErmon

Tweet media one

3

30

167

@baaadas

Jiaming Song

7 years

Basically you can use all kinds of regularization to maximize MI between data and code - e.g. GAN, Stein and MMD. Our experiments on PixelCNN show that MMD works the best, and can be implemented in 10 lines.

@StefanoErmon

Stefano Ermon

7 years

Check out our new blog post by @shengjia_zhao on InfoVAE:

Tweet media one

2

50

141

2

45

126

@baaadas

Jiaming Song

4 years

Can we make better use of negative samples in contrastive learning? In our #NeurIPS2020 paper, we show this is true by simply using a multi-label objective. Come to our oral presentation at 6:15 PT () and poster at 9-11 for more details! @StefanoErmon

Tweet media one

1

19

125

@baaadas

Jiaming Song

4 years

1. Iterative #generative models with Langevin dynamics-ish sampling can take a lot of time to sample, needing 1000s of iters. We show that they can actually get high quality samples already with only 10 - 20 steps!

2

18

119

@baaadas

Jiaming Song

9 months

Super proud to be part of the team! Join us, build multimodal foundation models, and unlock creativity for everyone:

Tweet card media

Luma AI - Careers

Make your imagination reality with AI.

@LumaLabsAI

Luma AI

9 months

🔥 Introducing Genie 1.0, our first step towards building multimodal AI. Genie is a text-to-3d model capable of creating any 3d object you can dream of in under 10 seconds with materials, quad mesh retopology, variable polycount, and in all standard formats! Try it on web and in

106

494

3K

7

4

113

@baaadas

Jiaming Song

4 years

International students are basically given 2 choices by #ICE . One: take physical courses and get infected by #COVID19 ; two: return to home country and possibly bring #COVID19 there. Wonder if ICE should be properly renamed as "International Coronavirus Emissary" 🤔

@ReichlinMelnick

Aaron Reichlin-Melnick

@ReichlinMelnick

4 years

This is bad. ICE just told students here on student visas that if their school is going online-only this fall, the students must depart the United States and cannot remain through the fall semester.

Tweet media one

3K

64K

80K

2

5

100

@baaadas

Jiaming Song

11 months

📢 Introducing Genie, a generative 3D foundation model that gives you 4x 3D models in 10 seconds! Try 🧞 on discord with: #MadeWithGenie

Tweet card media

Luma AI - Genie

A 3d generative foundation model

@LumaLabsAI

Luma AI

11 months

Meet Genie, a research preview of an all-new kind of generative 3D foundation model #MadeWithGenie 💬 Create 3D things in seconds on Discord ⚡️ Prototype in various styles 🎨 Customize materials 🆓 Free during research preview Try it now 👇

86

267

2K

4

7

97

@baaadas

Jiaming Song

11 months

Doing a Ph.D. in AI was sound back in 2016 where academia and industry tackle similar problem (e.g., CIFAR-10 generation). Now, unless you want to be a prof, why do a Ph.D.? At this rate of progress, is there even going to be much demand for AI research scientists in 5 years?

@sshkhr16

Shashank Shekhar

11 months

As PhD applications season draws closer, I have an alternative suggestion for people starting their careers in artificial intelligence/machine learning: Don't Do A PhD in Machine Learning ❌ (or, at least, not right now) 1/4 🧵

36

53

512

9

4

83

@baaadas

Jiaming Song

11 months

"MagicDiffusionDreamGAN is all you need for NeRFLLMs", a paper written by ChatGPT

@omerbartal

Omer Bar Tal

11 months

ICLR 2024 stats 💡 * 451 titles contain 'Diffusion' * 208 titles contain 'LLM' * 6 titles contain 'ChatGPT' * 25 titles contain 'NeRF' * 41 titles contain 'GAN' * 15 titles contain 'All you need' 🧐 * 22 titles contain 'Dream' * 6 titles contain 'Magic' #ICLR2024 @iclr_conf

Tweet media one

3

41

207

2

6

83

@baaadas

Jiaming Song

3 months

To be fair, the exact token count can be slightly different, as the tokener design in videos has not fully matured. But it is safe to say that the size (on disk) of the video datasets are hundreds times bigger than text datasets -- scaling can work quite differently here!

@AnjneyMidha

Anjney Midha 🇺🇸

3 months

Coming this week: the first ever research deep dive with @baaadas and @LumaLabsAI 's insane bet on 3D, and the new scaling curves it has unlocked

8

14

180

4

4

77

@baaadas

Jiaming Song

7 years

Got two papers accepted to #nips2017 ! InfoGAIL: A-NICE-MC:

Tweet card media

A-NICE-MC: Adversarial Training for MCMC

Existing Markov Chain Monte Carlo (MCMC) methods are either based on general-purpose and domain-agnostic schemes which can lead to slow convergence, or hand-crafting of problem-specific proposals...

3

12

75

@baaadas

Jiaming Song

4 months

Facts. Another thing I really like about working in a relatively small team (like ours) is that you never have to "fight" for compute resources or the opportunity to work on the "cool" stuff; instead, you can focus on getting real work done quicker and better. If that is

@hyperparticle

Dan Kondratyuk

4 months

@BenjaminDEKR Now you know why I left Google to join Luma. I was in the team that developed Veo early on but knew it would never be shipped to the masses for quite a long time, same as Sora. Not until a company like Luma forces their hand, that is (at least I hope, gimme access 😁).

29

41

483

3

3

76

@baaadas

Jiaming Song

11 months

The model definitions of consistency decoders are unfortunately not shared, so here is a guess from jit code. 1. Upsample the latent by 8x and concat with noised input (x0). 2. Run a few UNet down blocks (_1 to _16) and a midblock (_17). 3. A few UNet up blocks (_18 to _26).

Tweet media one

Tweet media two

@gabeeegoooh

Gabriel Goh

11 months

if you get the chance, do try decoding your Stable Diffusion generations with this decoder. You should see some improvements in text, small faces and straight lines. Made with @_tim_brooks , @DrYangSong , @model_mechanic , @txhf , @neonbjb

Tweet media one

5

34

193

1

4

74

@baaadas

Jiaming Song

6 years

HalfCheetah is an elusive environment. By simply reducing the variance initialization, one could obtain significantly higher sample complexity and a better local optima with the same algorithm (PPO). More details at

Tweet media one

2

10

62

@baaadas

Jiaming Song

10 months

This is the first year since 2017 when I have 0 accepted #NeurIPS papers... Nevertheless, could be nice to catch up on 12th and 13rd at #NOLA , hope to see you there! Btw we have a @LumaLabsAI party on the evening of 12th:

Tweet card media

Luma Happy Hour @ NeurIPS · Luma

Join us in NOLA for a fun evening with drinks, snacks and great conversations about all things generative AI!

1

3

62

@baaadas

Jiaming Song

6 years

Our UAI18 oral paper, A Lagrangian Perspective on Latent Variable Generative Models, is out Dual optimization for InfoVAE balances the trade-off between mutual information/ELBO/MMD, and Pareto dominates approaches with fixed trade-off hyperparameters

Tweet media one

0

19

60

@baaadas

Jiaming Song

10 months

Also, I am co-organizing the diffusion workshop with the amazing @ValentinDeBort1 @bahjat_kawar @_bunnech James Thornton @chenlin_meng and Jong Chul Ye. We have an amazing line up of speakers, poster, and panel discussion. Stay tuned!

@baaadas

Jiaming Song

10 months

This is the first year since 2017 when I have 0 accepted #NeurIPS papers... Nevertheless, could be nice to catch up on 12th and 13rd at #NOLA , hope to see you there! Btw we have a @LumaLabsAI party on the evening of 12th:

1

3

62

0

6

57

@baaadas

Jiaming Song

2 years

I am extremely honored to be called "the future of AI research" by @OpenAI 's ChatGPT!

Tweet media one

1

1

57

@baaadas

Jiaming Song

2 years

#ICML2022 Check out our talk at 5pm EDT on "Likelihood-free" Bayesian optimization! We are not presenting in person, but please shoot me a message if you have questions! Project page: @yulantao1996 @willieneis @StefanoErmon

Tweet media one

1

6

54

@baaadas

Jiaming Song

1 year

Just arrived at #SIGGRAPH2023 as a totally non-graphics person for the diffusion course (Thu 9-12, Rm 403AB). This covers some basics and some more recent developments, such as composing different concepts, controlling the generations, and 3d asset generation.

Tweet media one

1

6

51

@baaadas

Jiaming Song

8 months

Super impressive! "Similar to GPT models, Sora uses a transformer architecture, unlocking superior scaling performance." That might seems to suggest bigger is better is still true in diffusion models, unlike the conclusion in the Imagen paper that 3B is good enough...

@sama

Sam Altman

8 months

here is sora, our video generation model: today we are starting red-teaming and offering access to a limited number of creators. @_tim_brooks @billpeeb @model_mechanic are really incredible; amazing work by them and the team. remarkable moment.

2K

4K

26K

1

1

43

@baaadas

Jiaming Song

1 year

Not going to #ICML this year, but @qsh_zh will be presenting our paper on plug-and-play generation with diffusion models on Tuesday, Poster Session 2!

Tweet media one

1

9

43

@baaadas

Jiaming Song

6 years

After two hours of debugging, I found out that switching .Dataset.batch and affects performance significantly. On VAE on mnist, it amounts a 2x-3x slowdown if we put map before batch.

4

7

39

@baaadas

Jiaming Song

2 years

Come to Hall J 114 at 4p-6p to talk about our #NeurIPS work on how to apply pre-trained diffusion models for solving noisy inverse problems never seen during training, with 5-50x acceleration! Website: Joint with @bahjat_kawar Michael Elad @StefanoErmon

Tweet media one

0

5

39

@baaadas

Jiaming Song

1 year

🔥

Tweet media one

@baaadas

Jiaming Song

1 year

It is on now! Repeatedly clicking "admit all" on zoom for a while 🤣

0

1

18

0

1

39

@baaadas

Jiaming Song

6 years

Check out my poster on Multi-Agent Generative Adversarial Imitation Learning () from 5:00 - 7:00p, Room 517 #157 . @ermonste Also send me a pm if you want to chat about anything!

Tweet media one

0

11

39

@baaadas

Jiaming Song

1 year

🤩 Flythroughs from us @LumaLabsAI can turn your iPhone videos into cinematic walkthroughs, as if you are flying and looking through a FPV drone. 🛸 Create now with #3D #AI #NeRFs :

@LumaLabsAI

Luma AI

1 year

💫 Today we are incredibly excited to release Flythroughs, an app to show off your space with AI-generated cinematic videos that look like professional drone captures! No need for drones, lidar, expensive real estate cameras, and a crew. Record like you're showing the place to a

108

358

2K

0

4

39

@baaadas

Jiaming Song

6 years

My seq2seq model to me: "I'm not sure if you're going to have a great career, but if you're going to have a great career, you're going to have a great career." So true. What can I say. #AI_Philosophy

0

3

39

@baaadas

Jiaming Song

4 years

Experience replay with prioritization is critical to the success in Q-learning, but seems less so in actor-critic. In we introduce a plug-and-play prioritization approach to improve sample complexity of some actor-critic methods.

Tweet media one

3

7

38

@baaadas

Jiaming Song

5 years

Subject-wise acceptance rate of #AISTATS . Poor neural networks!

Tweet media one

1

4

37

@baaadas

Jiaming Song

3 years

Some recent works that aim to make diffusion models faster to sample from (for the #environment ): One trick is to use a deterministic path from latents to samples. Helpful if you want a sample in 50 steps.

Tweet card media

On Fast Sampling of Diffusion Probabilistic Models

In this work, we propose FastDPM, a unified framework for fast sampling in diffusion probabilistic models. FastDPM generalizes previous methods and gives rise to new algorithms with improved...

@s_scardapane

Simone Scardapane

3 years

*Score-based diffusion models* An emerging approach in generative modelling that is gathering more and more attention. If you are interested, I collected some introductive material and thoughts in a small thread. 👇 Feel free to weigh in with additional material! /n

Tweet media one

4

88

406

2

2

38

@baaadas

Jiaming Song

4 months

Thanks to @simonxxoo @ring_hyacinth for trying #LumaDreamMachine using images with strong styles. 🧵(1/n, images at the end of the thread).

5

5

38

@baaadas

Jiaming Song

5 years

Wait there is an Adam optimizer startup?

1

4

35

@baaadas

Jiaming Song

3 months

Going to #CVPR to give a talk on #LumaDreamMachine at the EDGE workshop (), from 11:30 to 12:10 today. Also lmk if you want to catch up! Video credit to @daken_ and @thomasneff93

2

1

37

@baaadas

Jiaming Song

11 months

Here is a tiger panda if you want to escape from @OpenAI discussions and hiring announcements in twitter for a moment.

Tweet media one

2

1

34

@baaadas

Jiaming Song

5 years

#NeurIPS workshop on Information Theory and Machine Learning starting at 920a, in E Hall A. The first workshop listed in the neurips book. We have an awesome line of speakers!

Tweet card media

Workshop Schedule 9:20-9:30 Introductory Remarks 9:30-10:00 [Invited Talk] Ayfer Özgür: Communication-Efficient Distributed Learning 10:00-10:40 [Invited Talk] Stefano Soatto and Alessandro Achille:...

sites.google.com

1

5

34

@baaadas

Jiaming Song

7 years

InfoGAIL w/ Yunzhu Li and @ermonste . We learn to distinguish and imitate human #selfdriving behaviors though end-to-end #DeepLearning RL. Poster #199 tonight. #NIPS2017

0

9

34

@baaadas

Jiaming Song

6 years

Scalable Natural Gradient Langevin Dynamics in Practice. Seems like SGLD and its variants (Riemaniann, K-FAC, RmsProp) *do not* perform better than vanilla SGD with a constant learning rate. "Data subsampling noise" rocks!

Tweet media one

Tweet media two

1

4

30

@baaadas

Jiaming Song

1 year

A confident military bear rendered in #GaussianSplatting . 🔥 Capture your favorite memories and interact with them with Luma AI @LumaLabsAI , here: #lumaai #3d #nerf #ai #aiguys

@LumaLabsAI

Luma AI

1 year

3D was either pretty, or fast. Now it’s BOTH! Meet Interactive Scenes built with Gaussian Splatting: 🔥Browser & Phone-Friendly: Hyperefficient and fast rendering everywhere 👌Embed Anywhere: 8-20MB streaming files (even smaller soon!) ✨Ultra High Quality offline NeRF renders &

57

335

2K

1

1

30

@baaadas

Jiaming Song

6 years

Current network architecture of the @OpenAI Dota agent, a lot of pooling and attentions!

Tweet media one

0

2

28

@baaadas

Jiaming Song

11 months

Agree with most part but not "unknown". DCGAN is the first work to show how to train GANs well, and the comes progressive GAN, Wasserstein GAN, Specteal Normalization GAN etc..

@jxmnop

jack morris

11 months

now seems as good a time as ever to remind people that the biggest breakthroughs at OpenAI came from a previously unknown researcher with a bachelors degree from olin college of engineering

Tweet media one

Tweet media two

Tweet media three

Tweet media four

63

526

6K

0

2

25

@baaadas

Jiaming Song

5 years

#NeurIPS2019 Workshop on Information Theory and Machine Learning is now open for submission! Deadline Sept 15 AOE.

Tweet card media

Workshop Schedule 9:20-9:30 Introductory Remarks 9:30-10:00 [Invited Talk] Ayfer Özgür: Communication-Efficient Distributed Learning 10:00-10:40 [Invited Talk] Stefano Soatto and Alessandro Achille:...

sites.google.com

1

3

25

@baaadas

Jiaming Song

6 months

Excited to work with you Dan!

@hyperparticle

Dan Kondratyuk

6 months

I'm excited to announce that I'll be joining @LumaLabsAI , working on some incredible stuff 😎

5

2

57

0

1

26

@baaadas

Jiaming Song

11 days

I am at the @PyTorch conference today! DM if you are interested in chatting about large-scale training and inference of diffusion and multimodal models with the @LumaLabsAI team!

@LumaLabsAI

Luma AI

13 days

🚀 Introducing the Dream Machine API. Developers can now build and scale creative products with the world's most popular and intuitive video generation model without building complex tools in their apps. Start today #LumaDreamMachine

70

287

2K

0

0

25

@baaadas

Jiaming Song

4 months

Need more GPUs brrr #LumaDreamMachine

@hyperparticle

Dan Kondratyuk

4 months

Due to extremely high demand our servers are burning! 🔥🔥🔥

27

21

302

2

0

24

@baaadas

Jiaming Song

11 months

Not created by me, but looks pretty cute:

Tweet media one

@andrewpprice

Andrew Price

11 months

Artists, QT with your most popular piece

51

128

2K

0

1

24

@baaadas

Jiaming Song

1 year

🤣 Ticked all of the boxes

4

0

24

@baaadas

Jiaming Song

4 years

In generative models, a common paradigm is to learn from "positive" data and rely on inductive biases to generalize; this fails on important properties, such as # of objects. Negative Data Augmentation allows us to tell the model what we **don't want**. #ICLR2021

Tweet media one

Tweet media two

@a7b2_3

Abhishek Sinha

4 years

Happy to share that our work "Negative Data Augmentation" has been accepted at #ICLR2021 . ) Joint work with @kmr_ayush @baaadas @burak_uzkent @StefanoErmon

3

16

84

0

3

24

@baaadas

Jiaming Song

1 year

Last supper but with corgis. Large content with #diffusion models in parallel, but no mixture of experts like #GPT4 . Poster #187 . #CVPR2023

Tweet media one

Tweet media two

Tweet media three

0

2

23

@baaadas

Jiaming Song

2 years

It turns out that one can perform certain types of image-to-image translation with two separate diffusion models, one for each domain. Of course, this can be extended to conditional diffusion models as well 🐱 -> 🐯 #iclr2023

@chenlin_meng

Chenlin Meng

2 years

Excited to share our #ICLR2023 paper "Dual Diffusion Implicit Bridges for Image-to-Image Translation" (DDIB), an image-to-image translation approach with diffusion models 🐱 -> 🐯

Tweet media one

1

26

132

0

3

22

@baaadas

Jiaming Song

4 months

Huge thanks to AWS for making Dream Machine a reality (as well as the generated video below)! Read more about it on:

@SwamiSivasubram

Swami Sivasubramanian

@SwamiSivasubram

4 months

We heard a lot from customers that the volume of data, size of the models, and time required for training FMs exponentially increased the complexity of training a model. That’s why we built @Amazon SageMaker HyperPod—to help organizations remove the heavy lifting involved with

7

16

106

0

0

22

@baaadas

Jiaming Song

6 years

Honored to receive the @Qualcomm 2018 Innovation Fellowship this year w/ @shengjia_zhao !

Tweet card media

Qualcomm Innovation Fellowship 2020 US

We believe that research and development is the key to harnessing the power of imagination and to discovering new possibilities. We are excited to announce a new kind of Fellowship that promotes...

www.qualcomm.com

0

0

22

@baaadas

Jiaming Song

5 months

Also, we have a happy hour on Wednesday: . See you there!

Tweet card media

Luma AI Happy Hour @ ICLR · Luma

Join Luma AI for an open-bar Happy Hour during ICLR 2024! 🍸 Relax and unwind with us after a long day of 'attending workshops' and 'deep intellectual…

@baaadas

Jiaming Song

5 months

Attending #ICLR2024 ? You should chat with @dome_271 and @pabloppp . They have done some fantastic work on Würstchen

0

1

18

0

0

20

@baaadas

Jiaming Song

2 years

Also, I will be at #NeurIPS2022 from Nov 29 to Dec 3, presenting a few works on diffusion models, score matching and RL. Shoot me a note if you want to chat! It's the first NeurIPS that does not overlap my birthday 😃

@baaadas

Jiaming Song

2 years

Minimum requirements is to have at least one first-author top-tier ML publication. Applications should be enrolled in a PhD program. Please apply using the form below. We will reach out to you if your application is selected for the next steps.

1

1

15

0

0

20

@baaadas

Jiaming Song

4 months

By the way, we have a banger party at #CVPR on Tuesday evening, sign up here: We are also looking for fantastic people to build with us and expand human imagination and creativity -- please shoot me a note if you are going to CVPR!

Tweet card media

Luma AI Happy Hour @ CVPR · Luma

Join Luma AI for an open-bar Happy Hour during CVPR 2024! 🍸 Relax and unwind with us after a long day of 'attending workshops' and 'deep intellectual…

@baaadas

Jiaming Song

4 months

Extremely proud to be working on this with many amazing people @LumaLabsAI ! Generate a 5-second, 120 frames video in 120 seconds from text or images *now* on: Available to everyone. #LumaDreamMachine

17

17

190

0

0

18

@baaadas

Jiaming Song

5 months

Attending #ICLR2024 ? You should chat with @dome_271 and @pabloppp . They have done some fantastic work on Würstchen

Tweet card media

Würstchen: An Efficient Architecture for Large-Scale Text-to-Image...

We introduce Würstchen, a novel architecture for text-to-image synthesis that combines competitive performance with unprecedented cost-effectiveness for large-scale text-to-image diffusion...

@LumaLabsAI

Luma AI

5 months

Attending @iclr_conf next week? Come grab a drink and connect with us during our happy hour. 🔗 #ICLR2024

1

9

43

0

1

18

@baaadas

Jiaming Song

6 years

Whatever. At least I can finally ditch TF for PyTorch after wasting my life figuring out how to do data pipelines in TF twice.

3

0

18

@baaadas

Jiaming Song

4 months

@Ethan_smith_20 We had concrete score matching that was a generalization of both discrete and continious score matching: And then there are a few new works that works on language modeling like

Tweet card media

Concrete Score Matching: Generalized Score Matching for Discrete Data

Representing probability distributions by the gradient of their density functions has proven effective in modeling a wide range of continuous data modalities. However, this representation is not...

1

3

19

@baaadas

Jiaming Song

6 years

Reflection, Refraction, and Hamiltonian Monte Carlo: one of the papers that I find to be super easy to explain -- refraction is when your energy is high enough to overcome a boundary, and reflection is when your energy is not.

Tweet media one

1

4

17

@baaadas

Jiaming Song

7 years

if you use deep RL to optimize for "flying cars" that is probably what you would get

1

1

17

@baaadas

Jiaming Song

1 year

@somu5an @ArashVahdat @chenlin_meng Yes. The slides are already here:

Tweet card media

Denoising Diffusion-based Generative Modeling: Foundations and Applications

Tutorial on Diffusion Models in Conjunction with CVPR 2023

cvpr2023-tutorial-diffusion-models.github.io

1

0

18

@baaadas

Jiaming Song

7 years

There is no way I could do the experiments for the #NIPS2017 paper without WGAN and WGAN-GP... Regular GAN is too unstable

@soumithchintala

Soumith Chintala

@soumithchintala

7 years

With WGAN, we've put forward a theory and limited experiments than I'd like in retrospect, could've done better. However, furthering validation of the method, a lot of papers use Wasserstein formulation, including NVIDIA's Progressive GANs (287 citations so far since January)

4

6

73

0

1

18

@baaadas

Jiaming Song

1 year

It is on now! Repeatedly clicking "admit all" on zoom for a while 🤣

@baaadas

Jiaming Song

1 year

📢 Our #CVPR2023 tutorial on "Denoising Diffusion Models: A Generative Learning Big Bang" w/ @chenlin_meng and @ArashVahdat is happening tomorrow morning! 9:00 to 12:30, West 202-204. This is the year of big bang for diffusion models in CVPR!

Tweet media one

6

37

197

0

1

18

@baaadas

Jiaming Song

2 years

🎯 diffusion models 🎯 accelerating diffusion models 🎯 generating image, video, 3D data, text, graph, motion 🎯 modeling molecule, protein, genome, climate, functional data 🎯 differential eqs 🎯 inverse problems 🎯 neural net efficiency, vision transformers 🎯 related topics

1

4

16

@baaadas

Jiaming Song

7 years

Have been trying the overleaf v2 feature just now; so it seems the v2 editor is based on the sharelatex one, which is better imao

Tweet media one

0

2

14

@baaadas

Jiaming Song

5 years

Learning Controllable Fair Representations -- an information-theoretic perspective on learning representations with controllable fairness, with Ria Kalluri and @ermonste -- -- Come check out out poster at #AISTATS We96 15:50p-18:50p today!

Tweet media one

0

1

15

@baaadas

Jiaming Song

6 years

Presenting 5 papers w/ coauthors at the Deep Generative Modeling workshop (Hall A5) #ICML2018

2

1

16

@baaadas

Jiaming Song

2 years

Minimum requirements is to have at least one first-author top-tier ML publication. Applications should be enrolled in a PhD program. Please apply using the form below. We will reach out to you if your application is selected for the next steps.

1

1

15

@baaadas

Jiaming Song

1 year

Come work with us on #GenerativeAI and #3d !

@LumaLabsAI

Luma AI

1 year

Do you want spectacular 3D generative AI and capture on Android? Want to build it with the best 3D AI team in the world? Apply now 👉🏽 #LumaAI #GenerativeAI #3D #androiddev #Android

0

10

47

0

0

14

@baaadas

Jiaming Song

1 year

We have compiled an initial list (), from which we will choose around 60 key papers to highlight (each taking 3 mins) in the tutorial. If you think a paper / an idea that you like is not on the list, please submit a pull request!

Tweet card media

GitHub - cvpr2023-tutorial-diffusion-models/papers

Contribute to cvpr2023-tutorial-diffusion-models/papers development by creating an account on GitHub.

1

3

14

@baaadas

Jiaming Song

7 years

Flow models could be super useful for deep MCMC!

@daniellevy__

Daniel Levy

7 years

Our paper (with Matt Hoffman and @jaschasd ) on learning MCMC kernels parameterized by neural networks was accepted to ICLR. Up to 106x ESS and improved posterior sampling for deep generative models. Paper: Code: cc: @GoogleBrain

0

44

143

1

1

14

@baaadas

Jiaming Song

2 years

Glad to have been a part of it! We added a simple number theory problem, and there seems to be much room for improvement🤣Now let's think step by step?

Tweet media one

Tweet media two

@jaschasd

Jascha Sohl-Dickstein

2 years

After 2 years of work by 442 contributors across 132 institutions, I am thrilled to announce that the paper is now live: . BIG-bench consists of 204 diverse tasks to measure and extrapolate the capabilities of large language models.

Tweet media one

35

571

3K

1

1

13

@baaadas

Jiaming Song

10 months

🩷 condolences to the amazing people @OpenAI -- I am sure great people can do great things no matter where they are

Tweet media one

0

1

13

@baaadas

Jiaming Song

7 years

Flow models trained with adversarial methods seems to perform worse (in terms of test NLL and Inception score) than a simple Gaussian Mixture Model that memorizes training data... By @adityagrover_

0

2

12

@baaadas

Jiaming Song

6 years

Poster tonight at Hall B 70!

@DrYangSong

Yang Song

6 years

Our new paper @icmlconf shows that making optimization trajectories more invariant to reparameterizations is helpful for improving convergence speed of natural gradient methods. Paper: Code:

Tweet media one

Tweet media two

Tweet media three

0

9

27

0

3

12

@baaadas

Jiaming Song

4 months

Meme Machine #LumaDreamMachine

@blizaine

Blaine Brown 

4 months

Dream Machine from @LumaLabsAI really brings memes to life! A thread 🧵

87

423

3K

0

0

11

@baaadas

Jiaming Song

7 years

Updated A-NICE-MC for NIPS camera ready version:

Tweet card media

A-NICE-MC: Adversarial Training for MCMC

Existing Markov Chain Monte Carlo (MCMC) methods are either based on general-purpose and domain-agnostic schemes which can lead to slow convergence, or hand-crafting of problem-specific proposals...

0

2

11

@baaadas

Jiaming Song

7 years

The #NIPS2017 line is way longer than the #RickandMorty #szechuansauce one I went to! Tells you something about the #AI #ML "excitements".

Tweet media one

0

2

10

@baaadas

Jiaming Song

1 year

This course cannot be possible without an amazing set of presenters: @chenlin_meng @ShuangL13799063 @StefanoErmon @junyanz89 @karsten_kreis @TsungYiLinCV @chenhsuan_lin

0

0

10

@baaadas

Jiaming Song

11 months

So essentially, the model still looks like a regular conditional diffusion UNet with concat conditioning, except the concatenation is over nearest-neighbor upsampled latents (for an image of 4x32x32 this would be upsampled to 4x256x256, so x0 is 7x256x256 after concat).

1

0

10

@baaadas

Jiaming Song

5 years

A very nice blog post about NTKs!

@mldcmu

Machine Learning Dept. at Carnegie Mellon

5 years

Ultra-Wide Deep Nets and the Neural Tangent Kernel (NTK) by Simon Du @mldcmu , and Wei Hu @Princeton Learn more → #deeplearning #machinelearning #ML #research

0

12

53

0

2

8

@baaadas

Jiaming Song

6 years

Work with @shengjia_zhao and @ermonste . Code here (trust me, it does not only contain a readme with "code coming soon")

Tweet card media

GitHub - ermongroup/lagvae: Lagrangian VAE

Lagrangian VAE. Contribute to ermongroup/lagvae development by creating an account on GitHub.

@baaadas

Jiaming Song

6 years

Our UAI18 oral paper, A Lagrangian Perspective on Latent Variable Generative Models, is out Dual optimization for InfoVAE balances the trade-off between mutual information/ELBO/MMD, and Pareto dominates approaches with fixed trade-off hyperparameters

Tweet media one

0

19

60

1

3

10

@baaadas

Jiaming Song

11 months

@jm_alexia Alexia, like I said, I believe doing PhD *in AI* was a great idea a few years back (and I don't deny being lucky enough to have had the opportunities to do so). But times change and if I were 22 now, I don't think I will do it (or be qualified to enter the strongest programs 🤣)

1

0

9

@baaadas

Jiaming Song

7 years

I had to cite the KDD paper just a few days ago... Linear SVMs are amazing

@yisongyue

Yisong Yue

7 years

It's so easy for younger researchers to lose sight of the amazing accomplishments by the more senior researchers in the field. Here's a little reminder of one of the best in the business.

1

21

61

0

1

8

@baaadas

Jiaming Song

4 months

❤️💕 @dome_271

@LumaLabsAI

Luma AI

4 months

This is Dream Machine, our first generative text-to-video and image-to-video model. This video showcases some of the capabilities of #LumaDreamMachine that we're most proud of. Try Dream Machine for free today 👉

80

113

721

0

1

8

@baaadas

Jiaming Song

2 years

@poolio The inpainting case is identical to reconstruction guidance, but with some hyperparameter choices making it perform as well as the task-specific Palette.

Tweet card media

Pseudoinverse-Guided Diffusion Models for Inverse Problems

We introduce pseudoinverse guidance, an approach to solve inverse problems with generative diffusion models.

0

0

8

@baaadas

Jiaming Song

5 years

Marking my first time reviewing a submission that does not mention anything about the paper in the abstract or introduction...

0

0

7

@baaadas

Jiaming Song

11 months

@gravicle Cat-erpillar

1

0

8

@baaadas

Jiaming Song

7 years

PPL for #pytorch !

@Tkaraletsos

Theofanis Karaletsos

7 years

Like Bayesian Inference and #pytorch ? Try our PPL, Pyro.

2

68

189

0

0

8

@baaadas

Jiaming Song

7 years

A recent paper by @zhaosj12 , me and @ermonste - why stop at GAN regularization when you can have more, such as MMD and Stein!

@StatMLPapers

Stat.ML Papers

7 years

InfoVAE: Information Maximizing Variational Autoencoders. (arXiv:1706.02262v1 [cs.LG])

0

10

58

0

0

8

@baaadas

Jiaming Song

5 years

Also, blog post: by Ria!

@baaadas

Jiaming Song

5 years

Learning Controllable Fair Representations -- an information-theoretic perspective on learning representations with controllable fairness, with Ria Kalluri and @ermonste -- -- Come check out out poster at #AISTATS We96 15:50p-18:50p today!

Tweet media one

0

1

15

0

1

6

@baaadas

Jiaming Song

6 years

Closed access journals is so 20th centur, and a shameless exploitation of researchers. Surprised why other fields are not adapting to the open access -- in biology everyone seems to publish in NCS, despite bioxiv

@tdietterich

Thomas G. Dietterich

6 years

Several machine learning researchers have signed a statement regarding the upcoming launch of Nature Machine Intelligence. If you agree, I encourage you to sign this as well.

32

1K

2K

0

0

7

@baaadas

Jiaming Song

7 years

An additional point for training caveats in RL: normalizing observations and rewards. You can even hide this in the environment:

1

1

6

@baaadas

Jiaming Song

4 months

(6/n)

1

1

7