Schedule
Zoom link
Legend
π₯ = 15 minutes talk π¨ = 30 minutes talk
π» = Tutorial π€ = Panel Discussion
π = Lunch Break β = Coffee Break
Monday 11 July 2022
08:30 - Registration
09:00 - Welcome Address
09:15 - π¨ Tess Smidt, π¨ Mario Geiger, and π¨ Jigyasa Nigam: Common background and theory video
Introductions to equivariance, group representations, tensor products, and density-based representations
11:00 - β
11:30 - π¨ Boris Kozinsky: Equivariant models and active learning of interatomic energies for large-scale dynamics video
12:00 - π₯ Brandon Wood : Inductive Biases for Generalizable Catalysis Models video
12:15 - π₯ Lisanne Knijff : Machine learning inference of molecular dipole moment in liquid water video
12:30 - π
13:30 13:45 - π¨ Kristof SchΓΌtt: Unifying machine learning and quantum chemistry with deep neural networks video
14:00 14:15 - π₯ Dari Kimanius: Heterogeneous cryo-EM reconstruction with learned priors video
14:15 14:30 - π₯ Yunqi Shao: PiNet-chi: learning equivariant tensors with invariant predictions video
14:30 14:45 - π¨ Andrea Grisafi: An atom-centered learning model of the electron density video
I will present an equivariant kernel method for doing the regression of the electron density decomposed on an atom-centered spherical harmonics basis.
15:00 15:15 - β
15:30 15:45 - π₯ David Kovacs: Extending Atomic Cluster Expansion to Equivariance video
15:45 16:00 - π₯ Stefaan Hessmann: Crystal structure search accelerated by neural network ensembles video
16:00 16:15 - π¨ Soledad Villar: Equivariant machine learning via classical invariant theory video (beginning is missing)
In this talk we show that it is simple to parameterize universally approximating polynomial functions that are equivariant under standard actions by the Euclidean, Lorentz, and PoincarΓ© groups, at any dimensionality d. The key observation is that nonlinear O(d)-equivariant (and related-group-equivariant) functions can be universally expressed in terms of a lightweight collection of (dimensionless) scalars -- scalar products and scalar contractions of the scalar, vector, and tensor inputs. We complement our theory with numerical examples that show that the scalar-based method is simple, efficient, and scalable.
16:30 16:45 - π¨ Risi Kondor: Permutation equivariant neural networks video (starts only at minute 36)
17:00 17:15 - π₯ Yi-Lun Liao: Equiformer: Equivariant Graph Attention Transformer for 3D Atomistic Graphs (https://arxiv.org/abs/2206.11990 )
Tuesday 12 July 2022
09:00 - π» Tess Smidt, Mario Geiger, e3nn code tutorial
Colab notebooks: Spherical Harmonics, Tensor Product
10:00 - π¨ Taco Cohen, Beyond Equivariance: Natural Transformations for Graphs NNs and Causal Models video
10:30 - π₯ Yi-Lun Liao: Equiformer: Equivariant Graph Attention Transformer for 3D Atomistic Graphs video
https://arxiv.org/abs/2206.11990
10:45 - π¨ Maurice Weiler, Coordinate independent convolutional networks on Riemannian manifolds video (starts at minute 15)
https://arxiv.org/abs/2106.06020
11:15 - β Only 15 min!
11:30 - π₯ Clement Vignac: Equivariance in set and graph generation video
11:45 - π₯ Danush Venkatesh: Equivariant neural network for image segmentation video
12:00 - π¨ Michele Ceriotti, A unified theory of equivariant machine learning video
Classical and last-generation symmetry-adapted schemes for atomistic machine learning can be understood as different implementations of the same equivariant mathematical structure.
12:30 - π
13:30 - π¨ Ilyes Batatia, Higher Order Equivariant Message Passing Neural Networks for Chemistry and Physics video
14:00 - π₯ Hubert Beck : Effortless generation of training data sets for machine learning potentials using active learning video
14:15 14:20 - π₯ Jiayan Xu: Recent Applications of Machine Learning Potentials in Heterogeneous Catalysis video
14:30 14:40 - π¨ Francesco Cagnetta, Structure beyond symmetry: locality and compositionality in deep CNNs video
15:10 - β
15:30 15:40 - π₯ Michael Pun: Learning the Shape of the Protein Universe with 3D-equivariant Holographic Convolutional Neural Networks video
15:45 16:00 - π₯ Sam Schoenholz : JAX MD: A Framework for Differentiable Molecular Physics link to demo video
16:20 - π¨ Moritz ThΓΌrlemann, Learning Atomic Multipoles: Prediction of the Electrostatic Potential with Equivariant Graph Neural Networks video
16:50 - π€ How does (or should?) ML in the physical sciences differ from general-purpose ML (Michele Ceriotti, Kristof SchΓΌtt, Tess Smidt, Philippe Schwaller)
19:00 - Conference dinner @ Gina
Location of Gina Restaurant, very close to EPFL metro.
Wednesday 13 July 2022
09:00 - π» Simon Batzner and Alby Musaelian, NequIP code Tutorial hungry-hungry-nequippo video
10:00 - π» Ivan Diaz, Voxel Convolution applied to Brain MRI Segmentation Tasks link to colab notebook video
11:00 - β
11:30 - π¨ Robin Winter, Unsupervised Learning of Group Invariant and Equivariant Representations video
We propose a general unsupervised learning strategy in which the latent representation is separated in a group-invariant term and an equivariant group action component.
12:00 - π€ Representations of signals in space
12:30 - π
The hiking map can be found here. The buses leave EPFL at 1345 and will then leave the hiking spot at 1830 to come back to Lausanne so don't be late or you risk getting left behind!
Trip to Creux du Van
We have coaches organized leaving from EPFL at 13:45 going to Creux du Van where those who want can join for a hike or stay and enjoy board- and lawn games and in beautiful surroundings.
Hangout at the lake
when: 19:30 is the bus is on time
where: Parc Louis Bourget, probably somewhere around 46.518376, 6.589001
Thursday 14 July 2022
09:00 - π¨ Josh Rackers, Beyond the Black Box: The potential and problems of equivariant electron densities video
Electron densities make an excellent case for testing the properties of equivariant learning models. I will discuss how equivariance can help us discover new physics of intermolecular interactions and the challenges associated with these kinds of tensorial problems.
09:30 - π¨ Alby Musaelian, Allegro: strictly local equivariant deep learning without message passing video
10:00 - π» Max Veit, Learning with tensorial SOAP using the librascal code link to Colab notebook video
In this tutorial I will demonstrate the use of equivariant (Ξ»-SOAP) and many-body (NICE) extensions of the SOAP framework for simple, flexible, and transparent learning of tensorial properties.
10:45 - π₯ Aria Mansouri: Machine-learning-directed development of functional materials video
11:00 - β
11:30 - π₯ Sander Vandenhaute: Machine Learning Potentials for Activated Processes using Active Learning video
11:45 - π» Guillaume Fraux, Equistore: sharing atomistic equivariant data Link to tutorial video
Equistore is a library to represent all kind of data and metadata found in equivariant atomistic machine learning, allowing to pass data between different libraries in the ecosystem.
12:30 - π
13:30 - π₯ Jonas Lederer: Equivariant Machine Learning for Molecular Nanorobotics video
13:45 - π₯ Artur Toshev: Equivariance in Smoothed Particle Hydrodynamics video
14:00 - π¨ Martin Uhrin: Invertible, compact and symmetry aware descriptions of atomic environments video
14:30 - π₯ Paul Spiering: Equivariant learning electronic friction tensors video
14:45 - π₯ Kiarash Jamali: Automated model building in cryo-EM maps using equivariant graph neural networks video
15:00 - β
15:30 - π¨ Ben Blum Smith, Invariant theory as a basis for equivariant models video
Invariant theory is an old and well-developed branch of algebra that describes functions obeying specified symmetries. We discuss the use of invariant theory in building equivariant models.
16:00 - π₯ Julian Tachella: Equivariant Imaging: Learning to Reconstruct Images Without Ground-Truth video
16:15 16:20 - π₯ Felix Musil : Developing machine learned coarse grained models for proteins and nuclear quantum effects video
16:30 - π€ Generalization beyond E(3)
Closing Remarks