Debuggable Deep Networks: Sparse Linear Models (Part 1)

We show how fitting sparse linear models over learned deep feature representations can lead to more debuggable neural networks while remaining highly accurate.

Unadversarial Examples: Designing Objects for Robust Vision

We show how to design objects to help, rather than hurt, the performance of vision systems; the resulting objects improve performance on natural and distribution-shifted data.

Benchmarks for Subpopulation Shift

We develop a methodology for constructing large-scale benchmarks to assess the robustness of standard models to subpopulation shift.

Transfer Learning with Adversarially Robust Models

We find that adversarially robust neural networks are better for downstream transfer learning than standard networks, despite having lower accuracy.

Noise or Signal: The Role of Backgrounds in Image Classification

To what extent to state-of-the-art vision models depend on image backgrounds?

From ImageNet to Image Classification

We take a closer look at the ImageNet dataset and identify ways in which it deviates from the underlying object recognition task.

Identifying Statistical Bias in Dataset Replication

Statistical bias in dataset reproduction studies can lead to skewed outcomes and observations.

Robustness Beyond Security: Computer Vision Applications

An off-the-shelf robust classifier can be used to perform a range of computer vision tasks beyond classification.