Working on enabling programmers to effectively deploy machine learning in large-scale production systems, with a focus on reducing engineering effort and scope for errors. See technical details: https://arxiv.org/abs/2304.13033

Developed a fast verifier (two to three orders of magnitude faster than the state of the art) to evaluate the robustness of neural networks to adversarial examples. The computational speedup enabled:

1) Verification of properties on convolutional and residual networks with over 100,000 ReLUs

2) Computation (for the first time) of the exact adversarial accuracy of a classifier subject to bounded perturbations

Across all networks and datasets considered, we certified more samples than the state-of-the-art and found more adversarial examples than a strong first-order attack. Analysis of factors affecting verification time led to follow-on work creating a regularizer to further reduce verification times.

Reduced worst-case latency of a versioned datastore by 10× by optimizing costliest code paths, enabling efficient access to terabytes of data.