🎊 It has arrived 🎊, the 2nd edition of my "Deep Generative Modeling" book. It has 100 new pages, 3 new chapters (incl.
#LLMs
) and new sections. It covers all deep generative models that constitute the core of all
#GenerativeAI
techs!
Check it out:
💻
I am excited to announce that my book, "Deep Generative Modeling", is available online and in print (
@SpringerNature
):
Code used in the book is freely available online: (1/4)
New year, and new challenges! I've decided to do a series of blog posts on deep generative modeling. The first post is out:
I hope you'll enjoy it! The next one in ~2 weeks 🤓
The fourth blog post on deep generative modeling! This time about one of my favorite topics: Variational Auto-Encoders ▶️◀️. Proposed concurrently by
@dpkingma
@wellingmax
and
@DaniloJRezende
@shakir_za
D. Wierstra.
👇
Post:
Code:
I am extremely happy to announce that yesterday, after ~12mo of hard work, I handed in the second edition of my book "Deep Generative Modeling" () to the publisher. There are ~100 pages of new content! ✨
💡I will keep you updated!!💡
A new blog post is out! This time about Diffusion-based Deep Generative Modeling! Read about this fascinating research line and check out the code:
Post 📜:
Code 💻:
The 7th post in the series on deep generative modeling. This time about priors ⏺️ in VAEs ▶️◀️, a very important topic that is often neglected. Join me in the journey on standard Gaussians, VampPriors, flows and GTMs.
👇
Post:
Can we efficiently learn one model that generates and classifies (i.e., a joint model)? Yes, we can! In our recent paper, we proposed a new joint model that combines a diffusion model with a classifier by sharing the UNet architecture.
Link :
[1/5]
After the summer break, the deep generative modeling series is back! 💪 This time about hierarchical latent variables models (hierarchical VAEs).
Post 📜:
Code 💻:
The end of February brings the last blog post in the short series on score-based generative models. After score matching and SDEs, it is time for flow matching!
Link to the post:
Link to the code:
Exactly one year ago, after writing it for 9mo (yep, it's my baby!), my book on
#GenerativeAI
was out! It was downloaded over 26k, which is over 70 times per day!
It was an amazing year for
#GeneartiveAI
, and if you'd like to catch up with this topic, just get this book 🤓
I am excited to announce that my book, "Deep Generative Modeling", is available online and in print (
@SpringerNature
):
Code used in the book is freely available online: (1/4)
Today is a new day and a new blog post on score-based generative models. How do we use score matching to learn generative models defined by a stochastic/ordinary differential equation? Check this out!
Post:
Code:
Kudos to Dr.
@emiel_hoogeboom
for his amazing Ph.D. project defended today! A
#GenerativeAI
superstar in all aspects!
P.S. I have never heard so witty and lovely laudatio (given by
@wellingmax
) before.
I played around a bit with KANs (Kolmogorov-Arnold Nets) parameterizing a VAE on a tiny data (1797 8x8 images of handwritten digits). Here are the results (ELBOs):
MLP (~0.5M): 96.8
KAN (~4M): 97.2
MLP+KAN (~2.7M): 96.7
KAN+MLP (~1.6M): 97.5
KAN (enc) and MLP (dec): 96.9
After the score-based models, I decided to take a step back and cover the basics in a new blog post: Probabilistic modeling and Mixture Models. Additionally, I make a brief intro to ✨Probabilistic Circuits✨ Check:
📄Post:
🖥️ Code:
New year, new blog posts! Today, I start with the first post from a series on score-based generative models. Are you curious about score matching and how to implement it? Check this out!
Post:
Code:
My recent work on general invertible transformations for flow-based generative models with some preliminary experiments. I present how we can generalize coupling layers and reversible logic operators. Paper: Code:
A magnificent paper on hybrid models using well-known generalized linear models and recent developments on flow-based generative models by
@eric_nalisnick
,
@amatsukawa
,
@yeewhye
, D. Gorur &
@balajiln
: . This is why generative modeling is important!
A beautiful paper by Bin Dai & David Wipf "Diagnosing and Enhancing VAE Models" ()! One of best appendixes I've ever read. Must read for anyone interested in VAEs. Thank you
@tolstikhini
for pointing out this paper!
As promised, today a new blog post is out! This time, we start with the first class of deep generative models in the series: Deep Autoregressive Models. Additionally, I provide a code that you could play around with!
Post:
Code:
A beautiful quote from Shannon. I have an impression that the AI field goes against this piece of wisdom. Companies and unis expect students to publish papers. Postdocs and asst. profs are expected to deliver >1 papers a year. We arrive at >9k submissions per conf.
#time4change
I decided to refresh my website () and... start a new (short) series of blog posts! Today the kickoff post: "Trouble in paradise: Does it make sense to train latent variable models with variational inference?"
Post:
Today we start
@NeurIPSConf
2023 in amazing New Orleans! Morever, I am excited to announce that I will be a program chair of
@NeurIPSConf
2024 together w/ Cheng Zhang &
@ulrichpaquet
, and
@DaniCMBelg
as the general chair. Great days ahead 💯
OK, so if anyone wants to delve into diffusion-based models, there is a cool paper about it by
@calvinyluo
:
And (shameful self-promotion) you can always first take a look at my book
#deepgenerativemodeling
:
1/n The NeurIPS 1st round of reviewing is coming. Again, I see more and more comments on Twitter like "please don't use another 'theorem' if it makes no sense", unclear theory doesn't make you paper better" or "English is so poor, why did you submit the paper".
The paper "Consistency Regularization for Variational Auto-Encoders" by
@_sam_sinha_
and
@adjiboussodieng
is by far the neatest paper I've read in 2021! So clean and smart idea, so amazing execution! A highly recommended read:
I am extremely excited to announce an opening (Ph.D. candidate) in my group at
@TUeindhoven
:
Please apply here:
Join us in the fantastic journey of developing
#GenerativeAI
!
The evaluations of my course came in. Here are some cherry-picked opinions: 🍒
"Jakub is an amazing teacher, which made the course very fun"
"Jakub is passionate about the field, and it makes the lectures very enjoyable"
"The lecturer was very knowledgeable and entertaining"
🫳🎤
A neat idea on utilizing the accept/reject sampling as the prior in the VAE by
@MatthiasSBauer
and
@AndriyMnih
"Resampled Priors for Variational Autoencoders".
Once the whole NeurIPS dust is down, it's time to return to the series on deep generative modeling! This time about GANs, a class of models that made a lot of fuzz in the world of DGMs.
Post 📜:
Code 💻:
The deadline for
@NeurIPSConf
2024 has passed, but it doesn't mean we don't work hard to have another fantastic conference! Right now, we are in need for REVIEWERS. If you know someone suitable for this role, please fill in this form:
Today is the second blog post in the new series. This time: "Are we dil(ff)usional about the ELBO?!".
What are the consequences of using Gaussian diffusion in latent variable models?
Blog:
Post:
#deepgenerativemodeling
#diffusion
Two wonderful papers from
@oxcsml
(
@yeewhye
group): "Hierarchical Representations with Poincaré Variational Auto-Encoders" () and "Probabilistic symmetry and invariant neural networks" ()!
Our paper "Conditional Channel Gated Networks for Task-Aware Continual Learning" got accepted to CVPR 2020! An amazing work done by
@davideabati
with
@Qualcomm_Tech
(
@TiRune
,
@BabakEht
, me) and
@UNIMORE_univ
(S. Calderara & R. Cucchiara). Working with Davide was a pure pleasure!
Together with
@jesfrellsen
&
@pamattei
, we organize Generative Modeling Summer School
#GeMSS23
! Please check out our webpage and apply!
Deadline: April 10, 2023
Where: Copenhagen (Denmark)
When: June 26-30, 2023
Lecturers: see below👇
Our paper on Wavelet Nets got accepted to
@TmlrOrg
@TmlrPub
! I am extremely happy about it for several reasons: 1) we never gave up, 2) it is a nice idea imho, 3) it is one of 17 papers published (crazy!) by
@davidwromero
during his PhD! Congrats David, Erik and Mark 👏
Wavelet Networks: Scale-Translation Equivariant Learning From Raw Time-Series
David W. Romero, Erik J Bekkers, Jakub M. Tomczak, Mark Hoogendoorn.
Action editor: Jeremias Sulam.
#wavelet_networks
#wavelet
#transforms
Finally, the paper on self-supervised VAEs is published in
@Entropy_MDPI
🔥 A great work w/
@JohnGatop
on incorporating deterministic and non-learnable transformations (e.g., downscaling, edge detection) in VAEs to produce sharper images.
Paper 👉:
I feel delighted about "Attention-based Multi-instance Mixed Models" being accepted to
@aistats_conf
as an oral! It is our attention-based MIL w/
@MaxIlse
on steroids: with pre-trained embeddings, variational inference and SOTA incl. histopathology!
Paper:
Mixed Models with Multiple Instance Learning (MixMIL) received an Oral & Outstanding Student Paper award at
@aistats_conf
last week! 🏆
MixMIL enables accurate & interpretable patient label prediction from single-cell data by adding attention to GLMMs.
#singlecell
#MachineLearning
The NeurIPS dust is down and now it's time to... apply to the asst. prof. of ML position at
@VUamsterdam
🤓 Come to work in Amsterdam 🇳🇱, the city hub of AI 💻🤖, with many brilliant ppl. 👩🦱👦🏾👨🦲🧑🏿🦱
Details:
An amazing work done by
@yperugachidiaz
on developing invertible dense networks (i-DenseNets) with learnable concatenation, an extension of Residual Flows (Chen et al., 2019). Paper: Code coming soon 😉
I am extremely proud to share a work on Approximate Bayesian Computation for discrete data (w/ Ilze A. Auzina). Applications: learning QMR-DT Network, (toyish) binary neural nets, and Neural Architecture Search.
Paper:
Code:
For everyone who missed the talk or wants to re-watch it, the slides and the recording of the talk are available online ():
Slides:
Video:
Thank you
@TIIuae
for the invitation!
I am happy to give a talk about deep generative modeling at
@TIIuae
next week on March 8. I will discuss why DGM is a key to unlocking the potential of AI!
More info here:
Hope to see you there!
Our (
@MaxIlse
@wellingmax
) latest work on an attention-based deep multiple instance learning with application to images and histopathology scans is on arXiv now:
Our new paper on domain invariance using VAEs w/
@MaxIlse
@ChrLouizos
&
@wellingmax
. Very interesting results on disentanglement & domain adaptation and a real application to medical imaging. Simply be like a DIVA!
GenAI does not stop, so it is time for a new blog post. Since LLMs are everywhere, I decided to take a look at them, my curious readers. And a bonus: An implementation of a teenyGPT 🤖✨
📄Post:
🖥️Code:
The 2nd day of the best summer school on GenAI: Generative Modeling Summer School
#GeMSS24
. After the morning session on ARMs (convs & transformers) given by yours truly, now we have a session on normalizing flows by
@jesf
Aaaaaannnnnnddddd... now it's accepted to ICML! Congrats
@MaxIlse
and Patrick Forre and thank you for an amazing collaboration! It's a great reward for perseverance.
#NeverGiveUp
💪
3 rejections (UAI, NeurIPS, AAAI) and 6 months later, I updated the arxiv version of our paper: Selecting Data Augmentation for Simulating Interventions. We have more theory and propose a simple algo to select data augmentations for domain generalization!
Big thanks to the organizers of
#GenU2021
for the invitation! It was a great pleasure to be in
#Copenhagen
and discuss deep generative models with so brilliant people! If you missed my talk, here are the slides:
PDF 🖥️:
Compression of neural networks using L0 norm is not differentiable. Or it is?
@ChrLouizos
@wellingmax
and
@dpkingma
show how to do that! A neat paper on learning sparse neural networks using L0 regularization:
I just finished Deep Generative Modeling by
@jmtomczak
. Covering topics from VAEs to GANs to EBMs, it has changed the way I think about ML. This book also demonstrates the reasoning behind the pesky sqrt(2) voodoo that shows up in so many neural networks.
After 8 months of long coding nights ☕️ we finally officially release Pythae 🥳, a python library unifying generative autoencoder implementations including vaegan🥗, vqvae or RAEs.
🖥️ github repo:
👉paper:
"An Information-Theoretic Analysis of Deep Latent-Variable Models" () by A.Alemi,
@poolio
, I.Fischer, J.Dillon, R.Saurous & K.Murphy is my favorite paper from the last year. A great insight into representation learning by analysing the mutual information.
In the era of
#GenerativeAI
, it is time to talk about GenAI systems. What are the components of GenAI Systems (GenAISys)? Does the future of AI rely on GenAISys (e.g., agentic AI 🤖)?
If you are interested, I'm happy to share my thoughts with you: ✨
I've just heard a very sad information that Prof. Conway passed away... As a ML researcher, I decided to make a tribute to him in the only way I could think of.
Congratulations to Dr.
@EmilevanKrieken
for defending his fantastic PhD WITH HONORS (cum laude)! Only ~5% of PhDs are distinguished with cum laude in the NL. As a co-promotor, I couldn't be more proud 🥲
I'm a bit late to the party, but I'm super excited about our paper on Bayesian optimization for combinatorial problems accepted to
@NeurIPSConf
. A great collaboration between
@Qualcomm
and
@quvalab
(ChangYong Oh,
@egavves
,
@wellingmax
). + I was selected top 400 reviewers 😏
Are you looking for a PhD position?
Do you want to work on continual learning, deep generative modeling and many more?
Apply now:
If you want to work with me, pick up the project No. 10!
More info:
I am very happy to be among TMLR Expert Reviewers🥇, i.e., reviewers for
@TmlrOrg
who have done exemplary work in evaluating
@TmlrOrg
submissions. Thank you 🙏
The full list:
Recently, I have been getting more interested in
#GenerativeAISystems
(
#GenAISys
). I am convinced that a systems-based perspective on
#GenAI
is crucial to design more reliable, robust, and responsive
#GenAI
.
You can read my perspective here:
We push
#GenerativeAI
forward! We updated our paper on joint diffusion models, and now we have new SOTAs on image classification and image synthesis in the class of joint models!
Moreover, we provide the code (many thanks to
@kamilrdeja
):
Can we efficiently learn one model that generates and classifies (i.e., a joint model)? Yes, we can! In our recent paper, we proposed a new joint model that combines a diffusion model with a classifier by sharing the UNet architecture.
Link :
[1/5]
My brilliant MSc student
@AmandaIlze
has her first paper accepted! I am very happy that our paper on Approximate Bayesian Computation for discrete spaces got accepted to
@Entropy_MDPI
.
👇
Code:
Paper:
A new exciting work w/ amazing
@JohnGatop
and M. Stol on generating crisp images using VAEs. The trick is to use image downscaling as a variational posterior and then super-resolution.
@CIGroupVU
+
@BrainCreators
Paper:
Code: