Code

The github pages contain more information for each individual project. If you are interested in code that is currently not on github, please send me an email.

Datasets

The following pages contain new test sets for CIFAR-10, ImageNet and SQuAD.

• CIFAR-10.1

• ImageNetV2

• SQuAD-shifts

The CIFAR-10.1 and ImageNetV2 repositories also provide code for assembling the datasets.

I have also worked on a reconstruction of the UCI Adult dataset as part of the folktables project.

Robustness

Our large-scale robustness benchmark for ImageNet: https://modestyachts.github.io/imagenet-testbed/.

Together with the MadryLab, I have worked on the following baseline implementation for adversarial robustness:

• MNIST challenge

• CIFAR-10 challenge

• Spatial transformations (rotations / translations)

Nearest neighbor

FALCONN is a fast implementation of locality-sensitive hashing (LSH) for cosine similarity. Ilya Razenshteyn and I developed this library.

Structured sparsity

When working with structure beyond sparsity, the the algorithmic core is often to projection onto a structured sparse set. The following github repositories contain code for various projection operators:

• Tree sparsity

• Graph sparsity

• EMD sparsity

The algorithm underlying the graph sparsity code is a fast implementation of the Goemans-Williamson scheme for prize-collecting Steiner forest (PCSF) problems. Aleksander Lenail is maintaining a version of the above PCSF code for applications in computational biology.

Distribution learning

The following repository contains code for differentially private distribution learning: github.