Instance-wise algorithm configuration with GNNs
Winning NeurIPS competition track entry (student leaderboard), on the ML4CO configuration task. We compile a large dataset of the solver performance and train a novel graph neural network that learns to predict a good configuration for a specific instance.
GPU accelerated Navier-Stokes solver
Julia implementation of a 3D multi-XPU diffusion solver and a 2D XPU Navier-Stokes solver based on matrix-free geometric Multigrid. Received honorable mention at
A highly optimized implementation of the FastSLAM algorithm achieving 60x speedup by using AVX (assembly) instructions, data restructuring and extensive benchmarking.
Scale- and Viewpoint-Robust Local Features
Using multi-level convolutional activations, we increase the robustness of local feature descriptors using a pretrained ResNet backbone with an attention layer for keypoint selection, which we train directly on the quality of the keypoints.