CVPR 2024 Workshop

Event: CVPR 2024 Workshop · Duration: 383 min · ▶ Watch on YouTube

Abstract

This workshop explores the latest advancements in equivariant deep learning, a field that leverages geometric symmetries to build more robust and efficient AI models. Speakers delve into fundamental concepts, practical applications, and theoretical underpinnings of equivariance. Topics include novel architectures, improved generalization, and the integration of physics-based priors into deep learning models for tasks like weather forecasting and molecular property prediction. The workshop also addresses challenges and future directions, such as scaling equivariant models to larger problems and understanding the role of hierarchical equivariance in complex systems.

Speakers

  • Stefanos Pertigkiozoglou — Google Research
  • Erik Bekkers — University of Amsterdam
  • Leonidas Guibas — Stanford University
  • Nina Miolane — UC Santa Barbara

Talks (3)

  • 00:00:00 — Stefanos Pertigkiozoglou: How to get started with equivariant deep learning
    • This talk provides an introduction to equivariant deep learning, covering fundamental concepts, benefits, and practical applications.
  • 01:57:27 — Erik Bekkers: Neural Ideograms and Geometry-Grounded Representation Learning
    • This talk introduces Neural Ideograms, a novel approach to representation learning grounded in geometric principles, offering insights into how brains encode information.
  • 03:38:00 — Nina Miolane: Hierarchical Equivariance in Artificial and Natural Brains
    • This talk explores hierarchical equivariance in both artificial and natural brains, focusing on how geometric structures in the world are encoded and processed.

Key Takeaways

  • Equivariant deep learning leverages geometric symmetries for robust and efficient AI models.
  • The field is moving towards more general and expressive equivariant architectures, including those for continuous groups and homogeneous spaces.
  • Equivariant models show promising results in diverse applications like weather forecasting, molecular property prediction, and robotic manipulation.
  • Hierarchical equivariance and the concept of ‘neural ideograms’ offer new perspectives on how brains encode and process information.
  • Scaling equivariant models to larger problems and addressing challenges like non-uniform sampling and probabilistic modeling are key open questions.

Methods / Models / Datasets Mentioned

  • SIFT Features
  • Steerable Filters
  • Neocognitron
  • LeNet
  • SO(3) Steerable Convolutional Kernel
  • Tensor Field Network
  • Harmonic Networks
  • E3NN
  • EMLP by Finzi et al. 2021
  • DWSNets
  • PointNet
  • DGCNN
  • UNet
  • GraphCast
  • Pangu-Weather
  • FuXi
  • Aurora
  • NeuralGCM
  • Kendall Shape-VAE
  • DINOv2
  • InfoNCE loss
  • SAM masks
  • LCA (Locally Competitive Algorithm)
  • Olshausen and Field (1996) model
  • Laplace prior
  • JAX library

Topics

Equivariant Deep Learning · Geometric Symmetries · Representation Learning · Neural Ideograms · Hierarchical Equivariance · Weather Forecasting · Molecular Property Prediction · Computer Vision · Robotics · Physics-based Priors

Notes

Open for commentary — connections to other work, critiques, follow-up reading.