Post Series
Recent Posts
Approximating Latent Manifolds in Neural Networks via Vanishing Ideals
Our ICML25 paper Approximating Latent Manifolds in Neural Networks via Vanishing Ideals introduces a novel approach to understanding data representation in neural networks through vanishing ideals. To characterize the structure of latent spaces, we scale vanishing ideal algorithms to handle high-dimensional data. The resulting algebraic characterizations can be used to replace standard network layers with more compact and efficient polynomial layers, leading to significant parameter reduction and potential improvements in inference speed while maintaining competitive performance.
Neural Discovery in Mathematics: Do Machines Dream of Colored Planes?
Our ICML 2025 Oral Paper Neural Discovery in Mathematics: Do Machines Dream of Colored Planes? introduces a novel neural network approach to tackle the famous Hadwiger-Nelson problem and related geometric coloring challenges. We reformulate the combinatorial task as a continuous optimization problem, enabling neural networks to find probabilistic colorings. This led to discovering two new 6-colorings, marking the first progress in 30 years on a key variant involving different forbidden distances and significantly expanding the known solution range.
Capturing Temporal Dynamics in Tree Canopy Height
Our paper Capturing Temporal Dynamics in Large-Scale Canopy Tree Height Estimation was accepted to ICML 2025! In this work, we present a novel approach to generate large-scale, high-resolution canopy height maps over time. Using Sentinel-2 time series satellite data and GEDI LiDAR data as ground truth, we present the first 10m resolution temporal canopy height map of the European continent for the period 2019-2022. Our pipeline and the resulting temporal height map are publicly available, enabling comprehensive large-scale monitoring of forests.
Global-Scale Forest Height Estimation
Our paper Estimating Canopy Height at Scale was accepted to ICML 2024! In this work, we present a novel framework for global-scale forest height estimation. Using a deep learning approach that leverages large amounts of satellite data with only sparsely distributed ground-truth height measurements from NASA's GEDI mission, we achieve state-of-the-art accuracy with MAE/RMSE of 2.43m/4.73m overall, significantly outperforming existing approaches. The resulting height map facilitates ecological analyses at a global scale.