2

Adversarial robustness of sparse local Lipschitz predictors

We study the stability of sparse activation patterns of neural networks. An extension of local Lipschitzness that accounts for invariance in activation patterns is provably better for studying robust certification and generalization.

Ramchandran Muthukumar, Jeremias Sulam

Adversarial robustness of sparse local Lipschitz predictors

Randomized sketching algorithms for low-memory dynamic optimization

Inexact optimization algorithms built on low-rank approximations of the PDE solutions is sufficient to find optimal control vectors.

Ramchandran Muthukumar, Drew P. Kouri, Madeleine Udell

Randomized sketching algorithms for low-memory dynamic optimization