Aaron's Learning
A showcase of my work on Data Science, Machine Learning and Neural Networks
Geometric Deep Learning Study Group
We’re working on our way through “Geometric Deep Learning: Grids, Groups, Graphs, Geodesics, and Gauges” by Michael M. Bronstein, Joan Bruna, Taco Cohen, Petar Veličković We’ll use the chapters rather than the big pdf, and also presenting on other papers.
To join the study group, fill out this form
Session 1: August 11, at 11am EDT/3pm GMT/5pm CEDT
Session 2: September 1, at 11am EDT/3pm GMT/5pm CEDT
Reading: Introduction
Lecture: ICLR 2021 Keynote: GDL: The Erlangen Programme of ML by M Bronstein”
Additional References: Mathematical Foundations of Geometric Deep Learning by Borde and Bronstein Introduction to Geometric Deep Learning by Patrick Nicolas
Presentation Topic: Geometric Deep Learning Reveals the Spatio-Temporal Features of Microscopic Motion Aaron Presenting
Session 3: September 22, at 11am EDT/3pm GMT/5pm CEDT
Reading: Graphs
Lecture: Graph Neural Networks with Petar Velickovic
Tutorial: Intro to GNNs
Optional Lectures on GNNs: Deep learning on graphs: successes, challenges by Bronstein
Graph Convolutional Networks by Federico Barbero
Graph Attention Networks by Federico Barbero
Presentations:
Working through “Intro to GNNs” Tutorial, Part 1
Possible Project on TacticAI
Session 4: October 13, at 11am EDT/3pm GMT/5pm CEDT
Lecture: Graph Neural Networks for Geometric Graphs by Chaitanya Joshi
Tutorials: Intro to GNNs, Part 2
Molecular Property Prediction with GNNs, Parts 0 and 1
Presentations: Expressivity and Under-reaching Working through “Molecular Property Prediction with GNNs” tutorial Deciding on which project to work on: Global Football, NFL Big Data Bowl, or Fake News Detection
Session 5: November 3, at 11am EST/4pm GMT/5pm CEST
Readings (blogs by Bronstein and collaborators): Expressive power of graph neural networks and the Weisfeiler-Lehman test Beyond Weisfeiler-Lehman: using substructures for provably expressive graph neural networks [GNNs through the lens of Differential Geometry and Algebraic Topology] (https://medium.com/data-science/graph-neural-networks-through-the-lens-of-differential-geometry-and-algebraic-topology-3a7c3c22d5f) Over-squashing, Bottlenecks, and Graph Ricci curvature
Lectures: Discrete Ollivier-Ricci curvature for data visualization and analysis by Abigail Hickok Watch the first 10 minutes to understand discrete curvature
[Curvature and Over-Squashing in GNNs by Francesco Giovanni](https://www.youtube.com/watch?v=pJyFg9NF7LQ) You can skip the first 10 minutes and start at 10:00
Tutorials: Molecular Property Prediction with GNNs, Parts 3-5
Session 6: November 17, at 11am EST/4pm GMT/5pm CEST
Readings: Foundations of Supervised Learning
Foundations of Equivariant Learning
Lecture: Equivariant Networks by Taco Cohen
Tutorials: Group Equivariant Neural Networks
Session 7: December 1, at 11am EST/4pm GMT/5pm CEST
Readings: Foundations of Geometric Learning
Possible Presentation Topics: Valence Labs Graph Learning on Wednesdays
Previous projects:
-
Privacy-preserving Detection of Violence in Cameras Writing a privacy product specification for a privacy-preserving neural network for video analysis as part of my capstone for the OpenMined course, “Our Privacy Opportunity.” Privacy-Preserving Violence Detector
- Mean Shift Analysis of the Uniform Crime Reporting data published by the FBI. I use TensorFlow 2.0 to do the analysis and bokeh to create the map.
-
Better Bin Packing Through Matrix Multiplication Using matrix multiplication to speed up a classic optimization problem
Excerpt of Metamorphosis II, by Marc Chagall
-
Optimizing Sailor Health for NavalX’s Hack The Machine. Part of second place finish in Data Science Track. Wrote code for Challenge 3: Optimizing Sailor Health
-
Pruned Neural Networks for Melanoma Detection Competing in Kaggle’s Melanoma Detection Competition through application of the Lottery Ticket Hypothesis for neural network pruning. Secondary goal of building more explainable artificial intelligence. Repo here
-
Quantum Computing for Network Analysis A proof-of-concept for identifying likely asymptomatic CoVID-19 carriers through network analysis by applying simulated quantum computing with qiskit. Entry into Qiskit Community Summer Jam 2020 https://qiskit-community-summer-jam-new-england.hackerearth.com/
End of Metamorphosis II, by Marc Chagall
Future plans:
-
Recursively defined sparse matrices A new way to build, add, and multiply sparse matrixes for graph neural networks and pruning