Romeo Valentin
- PhD Student at the Stanford Intelligent Systems Lab (SISL), developing certification guidelines for employing data-driven algorithms in safety-critical applications.
Sessions
I spent about one year leveraging Julia for DL research in Computer Vision, including using Flux.jl, FastAI.jl, Metalhead.jl, and loading python models through PyChainCall.jl. Happy to discuss my experience.
We present KSVD.jl, an extremely fast implementation of the K-SVD algorithm, including extensive single-core optimizations, shared-state multithreading, pipelined GPU offloading, and an optional distributed executor. With this implementation, we are able to outperform existing numpy-based implementations by ~100x and scale to datasets with millions of samples.
We present RunwayPNPSolve.jl, a framework for uncertainty-aware pose estimation for visual landing applications with multiple methods including real-time least-squares minimization + resampling, Monte-Carlo Markov-Chain, and a linear approximation, by leveraging the existing Julia package ecosystem.
The package further provides a framework of useful primitives to build simultaneously differentiable, unitful and coordinate-system-aware interfaces, and an interactive visualization pipeline.
We present ThreadedDenseSparseMul.jl, a library that efficiently computes dense-sparse multiplications and outperform competing packages (and Base.SparseArrays) in about 20 lines of code (for the basic functionality) by leveraging Polyester.jl. We further discuss the effect of Julia's memory layout on the performance and analyze the influence of different threading models.