About

I’m Simon Boehm.

Past Projects

• I built and maintain the fastest open-source decision tree compiler. It’s a fairly simple LLVM-based compiler. The speedups come from careful instruction cache blocking and lots of smaller tweaks (like bitvector-based encoding of categorical decision nodes). It works well enough, but should really be an interpreter instead (to help with the the icache bottleneck) so I’ll rewrite it eventually.
• I reimplemented many important algorithms for distributed deep learning (data parallelism, GPipe, interleaved backprop, …) to understand them better. The resulting library is called ShallowSpeed and runs on CPU using plain Numpy and MPI.
• I published a NeurIPS paper where we trained a deep learning model to predict the effects of drug molecules based on large single-cell RNA sequencing datasets.

FAQ

Drawings are made with Excalidraw.