AI Research at
@GoogleDeepMind
. Ex-Physicist. Graph Neural Networks & Controllable Generative Models (e.g. GCNs, Structured World Models, Slot Attention).
My PhD thesis "Deep Learning with Graph-Structured Representations" is now available for download: -- It covers a range of emerging topics in Deep Learning: from graph neural nets (and graph convolutions) to structure discovery (objects, relations, events)
Very excited about the release of Jraph. Finally an easy-to-use, extensive and fast Graph Neural Network library in JAX! Fully compatible with NN libraries such as Flax and Haiku:
I’m very excited to announce that I have joined the Google Brain team in Amsterdam 🚲🚣♀️🌳 as a Research Scientist! Looking fwd to both continuing my focus on graph-structured representation learning and to explore novel directions with an amazing team of new and old colleagues 😊
Excited to share our work
@GoogleAI
on Object-centric Learning with Slot Attention!
Slot Attention is a simple module for structure discovery and set prediction: it uses iterative attention to group perceptual inputs into a set of slots.
Paper:
[1/7]
So excited to share Object Scene Representation Transformer (OSRT):
OSRT learns about complex 3D scenes & decomposes them into objects w/o supervision, while rendering novel views up to 3000x faster than prior methods!
🖥️
📜
1/7
Excited to share our work on Contrastive Learning of Structured World Models!
C-SWMs learn object-factorized models & discover objects without supervision, using a simple loss inspired by work on graph embeddings
Paper:
Code:
1/5
Interested in the latest developments related to Graph Neural Nets and Structured Deep Learning? The talk recordings from the
@icmlconf
workshop on Learning and Reasoning with Graph-Structured Representations are now available:
Very honored to receive the ELLIS PhD Award for my thesis on Deep Learning with Graph-Structured Representations -- alongside with with
@NagraniArsha
for her work on multimodal DL (congrats!)
@ELLISforEurope
General Assembly happening just now, with ELLIS PhD Award 2021 to
@NagraniArsha
and
@thomaskipf
for very impressive theses on multimodal deep learning and graph NNs, respectively - congrats to both!
Life update: I've moved to the SF Bay Area! Excited to work more closely with my US-based
@GoogleDeepMind
colleagues and to meet both old and new friends in the area.
Leaving Amsterdam wasn't an easy decision: it's such an amazing city with a vibrant ML/AI community.
1/3
By far the cleanest and most elegant library for graph neural networks in PyTorch. Highly recommended! Unifies Capsule Nets (GNNs on bipartite graphs) and Transformers (GCNs with attention on fully-connected graphs) in a single API.
DGL (Deep Graph Library) - Clean and efficient library to build graph neural networks including GCN, TreeLSTM and graph generative models. Includes auto-batching and other tricks for speed.
"We [...] empirically show that explicitly enforcing roto-translation equivariance is not a strong requirement for generalization."
"Furthermore, we also show that approaches that do not explicitly enforce roto-translation equivariance (like ours) can match or outperform
1/n New preprint alert! Introducing Generative Molecular Conformer Fields (MCF) a generative model for molecular conformer generation that obtains state-of-the-art results without using any domain specific inductive biases!
Slides and recording of my talk/tutorial on Unsupervised Learning with Graph Neural Networks at the
@ipam_ucla
Workshop on Deep Geometric Learning of Big Data and Applications are now available:
Slides:
Recording:
Happy to share our new work on MolGAN: An implicit generative model for small molecular graphs. GANs + RL for generating graphs with desired properties: - Some benefits over VAEs, but can still suffer from mode collapse. Joint work with
@nicola_decao
Excited to share our work
@GoogleAI
on Object-centric Learning with Slot Attention!
Slot Attention is a simple module for structure discovery and set prediction: it uses iterative attention to group perceptual inputs into a set of slots.
Paper:
[1/7]
We just released a TensorFlow implementation of our Graph Convolutional Matrix Completion paper: (with
@vdbergrianne
) - A recommender system based on graph neural networks
Excited to share our work on Conditional Object-Centric Learning from Video!
We introduce SAVi, a slot-based model that can discover + represent visual entities in videos, using simple location cues and object motion (...or entirely unsupervised)
🖥️
1/7
Announcing AlphaFold 3: our state-of-the-art AI model for predicting the structure and interactions of all life’s molecules. 🧬
Here’s how we built it with
@IsomorphicLabs
and what it means for biology. 🧵
To showcase best practices for building/training Graph Neural Nets in JAX, we put together a comprehensive example for molecular activity prediction using Flax & Jraph
Official Flax GNN example:
Great work by
@BigAmeya
w/ collaborators @ Brain & DeepMind
"Neural Relational Inference for Interacting Systems" is accepted to
@icmlconf
(w/
@EthanFetaya
, Jackson Wang,
@wellingmax
, Rich Zemel)! Graph neural nets can learn to infer latent relational structure and model multi-agent and physical dynamics 👉
🚨NEW WEEKEND READING🚨
I'm publishing "Graph Neural Networks for Novice Math Fanatics" – a primer on the math behind GNNs using colourful drawings & diagrams
(I hope this becomes the definitive guide to GDL for those hoping to enter the field!!!)
PinSage: A New Graph Convolutional Neural Network for Web-Scale Recommender Systems - GCNs for network data made it from small academic benchmarks () to a web-scale industry application in less than two years time
Large-scale Graph Neural Network training without subsampling! Ma et al. (
@MSFTResearch
) implement a "Gather-Apply-Scatter" (GAS) approach in TensorFlow for Multi-GPU, distributed training of GNNs/GCNs:
The submission portal for the
@iclr_conf
Workshop on the Elements of Reasoning: Objects, Structure, and Causality (OSC) will open next Monday!
#osc2022
Submit your work by 25 Feb (end of day, AoE)
👉
Our implementation (in PyTorch) of "Neural Relational Inference for Interacting Systems" is available on GitHub: - includes a variant of Graph Neural Networks that learns both node- and edge-based representations. Joint work with
@EthanFetaya
We are organizing an
@ELLISforEurope
Workshop on Geometric and Relational Deep Learning! Registration invites will be shared soon. Interested in participating? Consider submitting an abstract or get in touch:
w/
@erikjbekkers
@wellingmax
@mmbronstein
Zero-shot Recognition via Semantic Embeddings and Knowledge Graphs (CVPR2018): Graph Convolutional Networks (GCNs) for zero-shot learning - massive improvements on a challenging ImageNet task
We are organizing an
@ELLISforEurope
Workshop on Geometric and Relational Deep Learning! Registration invites will be shared soon. Interested in participating? Consider submitting an abstract or get in touch:
w/
@erikjbekkers
@wellingmax
@mmbronstein
Excited to share our work on Contrastive Learning of Structured World Models!
C-SWMs learn object-factorized models & discover objects without supervision, using a simple loss inspired by work on graph embeddings
Paper:
Code:
1/5
Excited to announce DORSal: a 3D structured diffusion model for generation and object-level editing of 3D scenes.
DORSal is “geometry-free” and learns 3D scene structure purely from data – no expensive volume rendering!
🖥️
📜
1/6
Just finished reading through DeepMind's /
@PeterWBattaglia
et al.'s new review/position paper on graph neural nets: Timely and highly relevant contribution to the field IMO, couldn't agree more with their motivation for this class of models
Building Models that Learn to Discover Structure and Relations - A short post about our
@icmlconf
paper on Neural Relational Inference. w/
@EthanFetaya
Camera-ready versions of accepted papers at the
@NeurIPSConf
Graph Representation Learning Workshop 2019 are now online!
Some are missing, since camera-ready submission was optional (due to "no pre-print" policy of some other conferences/journals).
Our work on Hyperspherical Variational Auto-Encoders is accepted at
@uai2018
(plenary talk)! Great work from MSc students
@im_td
@lcfalors
@nicola_decao
(with
@jmtomczak
). Time to swap out your dusty Euclidean latent spaces 💨
Amsterdam is taking city infrastructure to the next level: can't stress enough how amazing these new, incredibly spacious (and surprisingly red) bike paths are.
Accepted papers at the
@icmlconf
Workshop on Learning and Reasoning with Graph-Structured Data are now available on the workshop website:
Papers:
Schedule:
New paper on learning hyperspherical latent spaces: Hyperspherical Variational Auto-Encoders (with
@im_td
, L. Falorsi,
@nicola_decao
,
@jmtomczak
). Useful trick for learning node embeddings in graphs and for semi-supervised learning.
Excited to share our work on Neural Assets: a new method for enabling 3D asset-level control in image diffusion models – scalable & without any 3D inductive biases.
Neural Assets goes beyond text or pixel-based control & provides an interface inspired by 3D graphics tools.
🧵
One year ago today, I defended my PhD thesis (). While the past year certainly has felt like any kind of celebration is inappropriate, I am still deeply grateful for everyone who helped or accompanied me along this journey.
Accepted papers are now available on the ICLR Objects, Structure & Causality (OSC) workshop website!
👉
We have received a total of 42 submissions out of which 31 will be presented at the workshop on April 29
We are organizing an
@ELLISforEurope
Workshop on Geometric and Relational Deep Learning! Registration invites will be shared soon. Interested in participating? Consider submitting an abstract or get in touch:
w/
@erikjbekkers
@wellingmax
@mmbronstein
Internship applications for Google Research are now open. If you are interested in working with some of us in the Brain Team in Europe (or elsewhere), please apply! Note that all internships will be virtual again next year.
It's exciting to see that Neural Relational Inference () can be re-purposed as a state-of-the-art causal discovery method (in a sense of non-linear Granger causality), entirely based on graph neural networks and variational inference.
New pre-print up with
@sindy_loewe
, Rich Zemel, and
@wellingmax
about causal discovery in time-series data!
"Amortized Causal Discovery: Learning to Infer Causal Graphs from Time-Series Data":
THREAD BELOW ⬇️⬇️⬇️
The list of accepted papers at the
@NeurIPSConf
Graph Representation Learning Workshop 2019 is online!
(Camera-ready versions will follow later this month).
Submission statistics / acceptance rates below 👇
Excited to share
@davide__belli
's work (with minor contributions by myself) on the Generative Graph Transformer, to be presented at the Graph Representation Learning workshop at
@NeurIPSConf
.
Blog:
Paper:
1/2
Graph Neural Networks for System Interaction Inference: A very nice article/tutorial by
@MJAljubran
on how to use Neural Relational Inference to infer hidden relations in interacting systems using GNNs
👉
SLATE (Slot Attention Transformer) is a new approach for unsupervised learning of compositional abstractions from visual input.
Combines Slot Attention w/ DALL-E to 1) bring object compositionality to DALL-E (in the absence of text prompts) and ...
1/2
Check out our new paper: "Illiterate DALL.E learns to compose." We propose a slot-based autoencoding arch, called SLATE, for learning object-centric representations that allow systematic generalization in zero-shot image generation without text!
Graph neural nets make their way into computer vision: - learn interaction graph (data-dependent) and propagate graph NN-style messages. Cool application!
Excited to share our work on self-supervised video object representation learning:
We introduce SAVi++, a slot-based video model that — for the first time — scales to Waymo Open driving scenes w/o direct supervision.
🖥️
📜
1/7
Excited (and honored) to be elected as an
@ELLISforEurope
Scholar in the newly launched "Semantic, Symbolic and Interpretable Machine Learning" ELLIS program.
Many thanks to
@vtresp
,
@kerstingAIML
& Paolo Frasconi for the nomination and to many others in ELLIS for their support!
Very happy to learn that our workshop proposal on Graph Representation Learning at
@NeurIPSConf
was accepted! w/ great co-organizers (see below) +
@vdbergrianne
,
@lrjconan
, Stefanie Jegelka and Yizhou Sun
New hot pre-print 🔥Interpreting Graph Neural Networks for NLP With Differentiable Edge Masking🔥
We show you can learn to remove most of the edges in GNNs such that the remaning ones are interpretable!
with
@michael_sejr
@iatitov
Looking forward to presenting our work on Slot Attention for Video (SAVi) at
@iclr_conf
!
Slot-based neural architectures (think of them as latent GNNs) are one the latest developments in ML I am most excited about. Lots of great submissions in this area at ICLR this year!
Excited to share our work on Conditional Object-Centric Learning from Video!
We introduce SAVi, a slot-based model that can discover + represent visual entities in videos, using simple location cues and object motion (...or entirely unsupervised)
🖥️
1/7
Our latest work at
@GoogleAI
on emergent object-centric learning will be presented at
@NeurIPSConf
this year:
1) OSRT: Object Scene Representation Transformer
So excited to share Object Scene Representation Transformer (OSRT):
OSRT learns about complex 3D scenes & decomposes them into objects w/o supervision, while rendering novel views up to 3000x faster than prior methods!
🖥️
📜
1/7
Proud to share our 150-page "proto-book" with
@mmbronstein
@joanbruna
@TacoCohen
on geometric DL! Through the lens of symmetries and invariances, we attempt to distill "all you need to build the architectures that are all you need".
More info below! 🧵
Want to learn more about object discovery, contrastive learning and world models? Visit our virtual poster at
#ICLR2020
(Thurs, Sessions 1&2) on Contrastive Learning with Structured World Models, w/
@ElisevanderPol
Happy to share our
#ICML2023
paper on Invariant Slot Attention (ISA)!
ISA learns per-slot reference frames, enabling pose control of objects & more data efficient learning.
📜
💻
🖼️
1/2
Most comfortable conference experience so far! Watching
@iclr_conf
#africaNLP
with
@maartjeterhoeve
(we both have a conference registration 🙃) from the comfort of our living room. Huge thanks to the conference/workshop organizers for making this happen!
Excited to share SlotFormer!
SlotFormer provides a simple recipe for visual simulation from pixels:
1) Learn a slot-based representation of a video
2) Simulate slot dynamics using a Transformer
(Led by
@Dazitu_616
, to be presented at ICLR 2023)
1/6
Happy to announce a Student Researcher opportunity in our team
@GoogleDeepMind
in Mountain View, California ☀️
If you’re excited about foundational research on controllability of visual/3D generative models, please consider applying:
CompILE discovers composable segments & encodings of behavior from sequential data (unsupervised and differentiable). Codes can be recomposed to facilitate generalization and exploration in RL.
New work with collaborators from
@DeepMindAI
Paper:
CompILE discovers composable segments & encodings of behavior from sequential data (unsupervised and differentiable). Codes can be recomposed to facilitate generalization and exploration in RL.
New work with collaborators from
@DeepMindAI
Paper:
I will be talking about Relational Structure Discovery today at the Graph Representation Learning (and Beyond) workshop at ICML. The talk will be live-streamed at 9:30 CEST. Join us (w/
@xbresson
) for live Q&A after the talk! Slide credit:
@PetarV_93
All you need is a graph neural net (for learning generalizable policies for structured agents). This is an interesting & exciting iteration on NerveNet () and its follow-up works.
RL gets specific to the robot it is trained on. Can a policy be trained to control many agents?
Turns out, training (shared) policy for each motor instead of whole robot not only achieves SOTA at train but also transfers to unseen agents w/o fine-tuning!
Accepted papers at the ELLIS Workshop on Geometric and Relational Deep Learning are now available on the workshop website: (incl. spotlight video and code for some)
We are organizing an
@ELLISforEurope
Workshop on Geometric and Relational Deep Learning! Registration invites will be shared soon. Interested in participating? Consider submitting an abstract or get in touch:
w/
@erikjbekkers
@wellingmax
@mmbronstein
xLSTM: Extended Long Short-Term Memory
The famous LSTM nets are improved by exponential gates and matrix-memory with covariance update. Strong results on large-scale language modeling.
P:
Excited to announce RUST (Really Unposed SRT):
RUST learns generalizable scene representations from raw images: no camera poses needed! A latent pose allows us to control the camera at test time 🎥
🖥️
📜
CVPR'23 (Highlight)
1/7
Couldn't agree more: we still don't have good solutions for scalable abstraction/concept learning from "raw" data (be it language, vision, or other sensory data). Solving this would likely unlock important new capabilities.
For anyone interested in future LLM development
One of the bigger unsolved deep learning problems: learning of hierarchical structure
Example: we still use tokenizers to train SOTA LLMs. We should be able to feed in bits/chars/bytes and get SOTA
Related: larger context window
This development is highly concerning, and makes it extremely difficult for candidates from non-CS backgrounds to get into these programs. I myself was lucky enough to get into a ML PhD program two years ago although my background was in Physics. Not sure if still possible today.
(Emergent) semantic segmentation from image-text pairs using Slot Attention
Jilan Xu,
@WeidiXie
et al.: Learning Open-vocabulary Semantic Segmentation Models From Natural Language Supervision
🖥️
📜