Quentin Oschatz

With circuit complexity being a major bottleneck in actually executing developed quantum algorithms in domains like protein modelling, we aim to leverage machine learning to optimize developed circuits without affecting correctness. By optimizing only small parts of the circuit at a time, we can verify that the substitution is either identical or within certain error bounds of the original sequence, allowing ML to be applied while trusting the results. This work is still in its early stages, with many avenues, architectures, and approaches still being explored.

Cerebus aims to verify the correctness of scientific code, especially when generated by LLMs. Here, mere syntactic correctness is insufficient, as scientific code can easily be plagued with hard-to-diagnose numerical errors despite compiling fine. Leveraging symbolic execution and semantic lifting, we can construct chains-of-reasoning that verify the numerical properties and error bounds of an implementation compared to the mathematical specifications.

This DARPA project aims to improve the basic scaling factor behind the processing of certain radar data, specifically when performing beamforming. Here, the clasical O(n³) Minimum Variance Distortionless Response (MVDR) beamforming algorithm is to be improved, ideally to yield a O(n log n) algorithm.

I helped develop and test several ML-based approaches, specifically CNNs acting on complex numbers. Additionally, I built a large part of the radar infrastructure, and helped implement a novel solver algorithm that took advantage of the data structure to reduce the scaling factor to O(n log n).