Invited talk

Principles for Initialization and Architecture Selection in Graph Neural Networks

SIMODS Minisymposium · SIAM Conference on Mathematics of Data Science (MDS26) · Salt Lake City · November 19, 2026

Three ways a deep GNN can fail before training starts — exponential outputs, oversmoothing, and correlation collapse — and the initialization and architecture choices that fix each one. Joint work with Boris Hanin.

  1. Slide 1 of 16
  2. Slide 2 of 16
  3. Slide 3 of 16
  4. Slide 4 of 16
  5. Slide 5 of 16
  6. Slide 6 of 16
  7. Slide 7 of 16
  8. Slide 8 of 16
  9. Slide 9 of 16
  10. Slide 10 of 16
  11. Slide 11 of 16
  12. Slide 12 of 16
  13. Slide 13 of 16
  14. Slide 14 of 16
  15. Slide 15 of 16
  16. Slide 16 of 16