📢🔥 My team at NVIDIA Research is looking for full-time research scientists & summer interns.
Topics of interest are:
1⃣Gen AI for science (climate, biology, chemistry)
2⃣Image/Video/3D (generate, edit, manipulate)
3⃣Fundamental generative learning
Apply via links below
I can't believe after this many years of programming with NumPy/PyTorch/TensorFlow, I didn't know about 𝚎𝚒𝚗𝚜𝚞𝚖. You can get rid of so many lines of reshape, transpose, sum, product & expand_dim with a single einsum which is even easier to understand:
📢📢📢 Introducing NVAE 📢📢📢
We show that deep hierarchical VAEs w/ carefully designed network architecture, generate high-quality images & achieve SOTA likelihood, even when trained w/ original VAE loss.
paper:
with
@jankautz
at
@NVIDIAAI
(1/n)
It breaks my 💚 when researchers tell me that VAEs don't work. My first typical question is "did you try hierarchial VAE or vanilla VAE?", the answer is usually vanilla VAE.
VAEs work much better with hierarchical structures. NVAEs and this work take this to the extreme!
Very Deep VAEs Generalize Autoregressive Models and Can Outperform Them on Images
“Very Deep VAEs” achieve higher likelihoods, use fewer parameters, generate samples 1000x faster, and are more easily applied to hi-res images, compared to PixelCNN.
📢📢 The recording of our tutorial on denoising diffusion models is now available on YouTube:
Slides and additional info:
This tutorial was originally presented by
@RuiqiGao
,
@karsten_kreis
, and myself at
#CVPR2022
.
🎉 Our tutorial proposal on 𝗗𝗲𝗻𝗼𝗶𝘀𝗶𝗻𝗴 𝗗𝗶𝗳𝗳𝘂𝘀𝗶𝗼𝗻-𝗯𝗮𝘀𝗲𝗱 𝗚𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗠𝗼𝗱𝗲𝗹𝗶𝗻𝗴 has been accepted to
#CVPR2022
. Stay tuned for some cool lectures from
@karsten_kreis
,
@RuiqiGao
and me at
@CVPR
2022.
This is me while explaining relaxed Boltzmann machines and not knowing that the pioneer who introduced them to the field is just behind me :(
#NeurIPS2018
📢 Would you like to build latent diffusion models for new problem domains? Wondering about theoretical and practical considerations!
@RuiqiGao
,
@karsten_kreis
and I will present the
#NeurIPS2023
tutorial on "Latent Diffusion Models" on Monday, Dec 11.
📢🔥 My team at NVIDIA research is looking for candidates with a fundamental generative learning background (ideally) in one of these domains:
- Gen AI for science (climate, chemistry, biology)
- Gen AI for 3D data
Apply via:
📢🔥My team at NVIDIA Research is looking for Summer 2024 interns. Topics of interest are:
1⃣ Fundamental generative learning (diffusion, etc)
2⃣Gen AI for science (climate, biology, chemistry)
3⃣Image/Video/3D (generate, edit, manipulate)
Apply via
📢 NVAE source code is released!
NVAE is deep hierarchical VAE w/ specially designed network architecture that can generate high-quality images & achieve SOTA log-likelihood.
Happy coding!
paper:
code:
w/
@jankautz
at
@NVIDIAAI
📢 I am looking for Fall interns to join the fundamental gen AI research team at NVResearch.
Topics of interest are:
-Fundamental Gen AI research
-3D generation
-Text-based manipulation
Remote/in-person, full/part-time options are available.
Apply at
Recent text-to-image models such as
#dalle2
or
#Imagen
are trained with ~800m image-text pairs. If that data was a movie (at 10 fps), it would take us 2.5 years to watch it all non-stop.
📢 Latent Score-based Generative Model (LSGM)
Score-based generative models (SGMs) are applied directly in data space & often require 1000s of network evaluations for sampling. We introduce LSGM for training SGMs in a latent space.
w/
@karsten_kreis
(eq. contrib.) &
@jankautz
🔥Denoising Diffusion GAN🔥
We tackle the 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝘁𝗿𝗶𝗹𝗲𝗺𝗺𝗮 with the novel 𝗱𝗲𝗻𝗼𝗶𝘀𝗶𝗻𝗴 𝗱𝗶𝗳𝗳𝘂𝘀𝗶𝗼𝗻 𝗚𝗔𝗡, a diffusion model designed for fast sampling.
Zhisheng Xiao
@karsten_kreis
ft. Peanut 🐈
📢
#LACE
: Controllable & Compositional Generation with Latent-Space Energy-Based Models
We introduce an extremely simple method for converting unconditional GANs into conditional models by training only a classifier.
abs:
project:
🔥 VAEBM = VAE + EBM 🔥
VAEBM is a new generative model defined by a symbiotic composition of VAE and EBM. VAEBM improves generative quality of VAEs & explicit EBMs by a large margin & reduces the gap with GANs.
w/ Zhisheng Xiao, Karsten Kreis,
@jankautz
At
#ICML2023
, I'm hiring Fall interns & research scientists for fundamental gen AI roles. If your background is relevant, shoot me an email with your CV.
Job postings:
📢 Diffusion Models for Adversarial Purification (
#ICML2022
)
Given an adversarial image, our DiffPure removes adv. perturbations by adding noise following the forward diffusion process, then denoising using the reverse generative process.
project page:
NVAE's accepted to
#NeurIPS2020
as a
#spotlight
paper! Many thanks to anonymous reviewers & AC who recognized the potential impact of NVAE
The code is already released here:
You'll soon hear about exciting stuff we've been doing. w/
@jankautz
@NVIDIAAI
📢📢📢 Introducing NVAE 📢📢📢
We show that deep hierarchical VAEs w/ carefully designed network architecture, generate high-quality images & achieve SOTA likelihood, even when trained w/ original VAE loss.
paper:
with
@jankautz
at
@NVIDIAAI
(1/n)
📢📢 NCP-VAE: VAEs w/ Noise Contrastive Priors
We tackle the prior hole problem: prior's failure to match the aggregate posterior. We train a reweighting term in prior using noise contrastive estimation.
w/
@JyotiAneja
@alexschwing
@jankautz
Likelihood-based generative learning is really really hard! Covering all the data distribution while having high sample quality and fast generation is probably one of the biggest challenges that the research community is facing currently.
Are you working on score-based models for:
1⃣Generative diffusion models
2⃣Differential eqn
3⃣Optimal transport
4⃣Stein’s method
5⃣Cool applications
Consider submitting your work to our
#NeurIPS2022
workshop on score-based methods:
Deadline: 22 Sept 2022
📢
#CVPR2023
Spontaneous Diffusion Meetup
Several folks have asked me to do a casual meetup for those interested in diffusion models
@CVPR
. Let's meet tomorrow (Wed) at 1 pm in West building (upper levels). Exact location will be posted below this tweet.
Please spread the word
📢 We are looking for candidates with a strong background in generative learning on text, image, video, graph & 3D data to join our team
@nvidia
. We strive for high-impact forward-looking research.
Apply via:
Please share!
The last 4-5 weeks have been stressful for me since my dad in his late 70s was diagnosed with
#COVID19
, and my mom, also in her 70s, refused to live separately away from my dad while he was going through this. My parents live in Iran, ~12,000 KMs away from where I am.
📢 Our team is hiring research scientists in various areas in machine learning & computer vision including 𝗮𝗱𝘃𝗮𝗻𝗰𝗲𝗱 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗺𝗼𝗱𝗲𝗹𝘀. Please apply below especially if you have worked on generative models such as diffusion models.
Discrete math is one of the reasons I hated computer engineering in undergrad.
I was determined to change my major to mechanical or electrical engineering because of their core math training. I changed my mind later after taking computer vision & signal processing courses.
Check our
#GTC23
session on "GANs to Diffusion - the path to
#generativeAI
" where Mitesh Patel and I talk about gen AI's past, present, and future. We discuss deep generative models, their shortcomings, and how diffusion models helped us solve them.
link:
Were you always suspicious of VAEs? 👀 Too blurry? Unstable? 😱 Well stop right there, because
@NVIDIAAI
has built NVAE, a hierarchical multi-scale VAE that can output crispy clear samples at high resolution 💃🔥 Watch the Video!
@ArashVahdat
@jankautz
🔥
#ICLR2021
#Spotlight
🔥
Our VAEBM has been accepted to ICLR as a spotlight.
VAEBM was results of an amazing collaboration with Zhisheng Xiao (University of Chicago), Karsten Kreis (NVIDIA's Toronto AI lab) &
@jankautz
(NVResearch Boston) spanning 4 locations & 3 timezones
🔥 VAEBM = VAE + EBM 🔥
VAEBM is a new generative model defined by a symbiotic composition of VAE and EBM. VAEBM improves generative quality of VAEs & explicit EBMs by a large margin & reduces the gap with GANs.
w/ Zhisheng Xiao, Karsten Kreis,
@jankautz
I am thrilled to announce that I have joined
@nvidia
research as a senior research scientist. I will continue tackling challenging problems in machine learning and computer vision along with brilliant
@NvidiaAI
researchers
@jankautz
@AnimaAnandkumar
@liu_mingyu
looking for an intern to work with me at NVIDIA research next summer on one of these topics:
- likelihood-based generative learning
- Probabilistic weakly-supervised models (e.g., noisy labels)
- Representation learning
- Discrete latent models
- Neural Architecture Search 👇
How can we learn vision tasks with noisy or no labels? What if we have only a pre-trained network and no data to learn from? Explore these and more exiting new frontiers for learning with limited labels or data in our upcoming
#ECCV2020
tutorial:
If you are a Ph.D. student, consider applying to NVIDIA's graduate fellowship program, especially if you're working on artificial intelligence, robotics, autonomous vehicles, and related fields.
What I was about to share:
Yeayy! I got free registration to NeurIPS as one of the top reviewers
What I wasn't going to share:
2 papers rejected. I forgot that AAAI deadline was last week
Morality:
Twitter shows the successful face of us. Never get fooled! Never give up!
There should be a process to penalize irresponsive reviewers. There are a couple of ICML reviewers in my batch that never bothered to submit their reviews and never replied to my emails/messages when asking about it. This happens at other conferences too.
Today, Jensen announced our work on CorrDiff in the
#GTC24
keynote. CorrDiff is a new generative model that superresolves weather events from 25km to 2km resolution, with 1,000x the speed and 3,000x the energy efficiency of conventional models.
At
#NeurIPS2022
, we are holding a workshop on score-based methods and diffusion models that will feature a panel discussion among experts in this space.
Help us ask them intriguing questions by filling out this form:
More info at:
The code for "Learning Undirected Posteriors by Backpropagation through MCMC" is released. I had lots of fun working on this. The paper comes with in-depth discussion of possible future works, ideal for summer interns😉
paper
code
Geoffrey Hinton is such a humble and humorous person! Finally, got the chance to hear his thoughts about our work on marrying Boltzmann machines and deep generative learning
#NeurIPS2018
If you are interested in combining energy-based models and VAEs, check out our
#ICML2020
paper that proposes an approach for training EBMs as approximate posteriors:
poster session: Wed 8 am and 9 pm PT.
Our work on training undirected graphical models (UGMs) as approximate posteriors is accepted to
#ICML2020
. We show that UGMs can be trained by backpropagation through MCMC steps. We examine these ideas on RBM approximate posteriors in VAEs
paper w/ code
Machine learning Twitter: how do you pass a large list of arguments to your python training scripts? If you are happy with any other library please comment below.
We have released the
#LSGM
source code with a brand new project page at:
If you're interested in denoising diffusion models (a.k.a. score-based generative models), come to our poster session next week at
#NeurIPS2021
.
It is really cool to see that denoising diffusion GANs have already been applied to speech synthesis just 4 months after our paper became available on OpenReview 🤯
@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"!
🔥Denoising Diffusion GAN🔥
We tackle the 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝘁𝗿𝗶𝗹𝗲𝗺𝗺𝗮 with the novel 𝗱𝗲𝗻𝗼𝗶𝘀𝗶𝗻𝗴 𝗱𝗶𝗳𝗳𝘂𝘀𝗶𝗼𝗻 𝗚𝗔𝗡, a diffusion model designed for fast sampling.
Zhisheng Xiao
@karsten_kreis
ft. Peanut 🐈
Check out our new
#cvpr2021
paper! We propose a simple technique to recover training images from gradients during training, questioning the efficacy of federated learning techniques in preserving data privacy.
Happy to introduce GradInversion
#CVPR2021
- gradients encode a surprisingly large amount of information, such that all the individual images can be completely recovered, even for complex ImageNet, deep networks, over large batch sizes.
Paper:
Did you know when you take the geometric mean of several distributions, you're minimizing a KL divergence?
In this post, I show how the geometric mean is related to the KL divergence, and how we can use this relation for training from noisy labels.
Post:
LSGM will be presented at
#NeurIPS2021
🎉
In this work, we show how score-based generative models (a.k.a. denoising diffusion models) can be trained in a latent space.
📢 Latent Score-based Generative Model (LSGM)
Score-based generative models (SGMs) are applied directly in data space & often require 1000s of network evaluations for sampling. We introduce LSGM for training SGMs in a latent space.
w/
@karsten_kreis
(eq. contrib.) &
@jankautz
Totally agree!
Anyone screening applications and any applicant thinking their CV is not representative of their skills/potentials, I think you might want to read the story of my own PhD application in this thread:
1/
📢 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!
.
@iclr_conf
this year allowed self-nomination for Area Chairs which gave me a chance to nominate myself and serve as a junior AC after 10 years of publishing at top conferences. It would be great if other conferences like
@NeurIPSConf
,
@icmlconf
,
@CVPR
did the same.
I was talking to an ECR (PI level) who has reviewed and published in major ML conferences for years, but never been invited as an Area Chair. I am wondering how to improve this.
Do others relate to this? Happy to hear your opinion (DM me if uncomfortable tweeting in public).
Yesterday we ran into several connection issues & we couldn't stream our
@CVPR
tutorial on diffusion models properly online.
I apologize for all the inconvenience. We are going to redo all our presentations and post them to the website in the coming weeks.
Stay tuned!
#CVPR2022
My road bike was stolen a few months ago, so I got a new Trek Domane. This bike is one of the popular bikes used in UNPAVED segments of Tour De France, which makes it a perfect bike for PAVED roads in the bay area 🤦♂️
I am very happy that my family managed to beat covid, and did so without transmitting it to each other so far. Folks, take this disease seriously! I hope none of you and your loved ones go through
#COVID19
.
Our score-based generative modeling with critically-damped Langevin diffusion paper will be presented at
#ICLR2022
as a spotlight paper! 🎉
Amazing work with
@timudk
&
@karsten_kreis
!
📢 Score-based Generative Modeling with Critically-Damped Langevin Diffusion!
We propose a novel diffusion using auxiliary velocity variables for more efficient denoising and higher quality generative models.
w/ the amazing
@timudk
&
@ArashVahdat
!
(1/n)
📢 HANT: Hardware-Aware Network Transformation
How can we accelerate a trained network to meet efficiency requirements for deployment? The current network compression methods such as pruning, kernel fusion & quantization address this w/o changing underlying network operations.
Alias-Free GAN
pdf:
project page:
networks match the FID of StyleGAN2 but differ dramatically in their internal representations, and
they are fully equivariant to translation and rotation even at subpixel scales
Introducing "NVAE", the first successful VAE applied to natural images as large as 256×256 pixels!
@ArashVahdat
@JanKautz
Watch the full-length demo of NVAE here:
#NeurIPS2020
So this week, the
#ML
community is attending
#ICLR2020
, while discussing rebuttals for the next conference
#ICML2020
, preparing their drafts for the next next conference
#NeurIPS2020
and wondering how to slow down.
I'm attending
#GTC24
for the first time in person, and I'm genuinely impressed by the preparation that the green team has put into this. Well done
@nvidia
!
Sarina Esmailzadeh, 16, beaten to death by Iranian security forces for protesting for women’s rights in Iran. Nika Shakarami, 16, killed after burning her headscarf in protest. Hadis Najafi, 23, shot multiple times during demonstrations sparked by the death of Mahsa Amini, 22,
Happen to have experience with neural architecture search?
👉 Send me your CV if you are interested in working on NAS as a summer intern at
@NvidiaAI
research.
👉 Burned out with
#ICML2020
and
#CVPR2020
deadlines? Get back to me only when you are well-rested!!
It turned out that the
#NeurIPS2021
supplementary material deadline can be more stressful than the main conference deadline when you have an ∞ number of pages in the appendices.
Confirmed!
#NeurIPS2021
reviews will be done on openreview! This is great change that hopefully will encourage more discussion during the review period!
Our work on training undirected graphical models (UGMs) as approximate posteriors is accepted to
#ICML2020
. We show that UGMs can be trained by backpropagation through MCMC steps. We examine these ideas on RBM approximate posteriors in VAEs
paper w/ code
@rasbt
IMO they're a bit hard to pick up requiring working knowledge of SDEs/ODEs/VAE/norm. flows.
They're very easy & stable to train, but also slow to sample from.
These two ICLR papers have reduced their sampling cost dramatically:
📢 Denoising diffusion GAN has been accepted to
#ICLR2022
as a spotlight paper 🎉
w/ the amazing Zhisheng Xiao &
@karsten_kreis
📢 I will present this work on Friday at
@ml_collective
reading group organized by
@savvyRL
. Sign up if you're interested:
🔥Denoising Diffusion GAN🔥
We tackle the 𝗴𝗲𝗻𝗲𝗿𝗮𝘁𝗶𝘃𝗲 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝘁𝗿𝗶𝗹𝗲𝗺𝗺𝗮 with the novel 𝗱𝗲𝗻𝗼𝗶𝘀𝗶𝗻𝗴 𝗱𝗶𝗳𝗳𝘂𝘀𝗶𝗼𝗻 𝗚𝗔𝗡, a diffusion model designed for fast sampling.
Zhisheng Xiao
@karsten_kreis
ft. Peanut 🐈
Iranian women are cutting their hair in front of the camera to protest the death of
#Mahsa_Amini
who was killed during a "morality" police arrest in Iran for not having a "proper" hijab. 💔