Research Landscape
Each paper is encoded as a vector of research concepts (e.g. bayesian inference, equivariant neural networks), then projected to 2D so that papers with similar concept profiles land nearby. Hover any paper to reveal which others share its concepts (highlighted in purple).
UMAP — preserves both local clusters and the distances between them. The most faithful overall map.
t-SNE — tight, well-separated clusters. But cluster positions relative to each other are arbitrary — don't read meaning into which cluster is "left" or "right".
PCA — linear projection. The horizontal axis is roughly the direction of maximum variance across all concept vectors; the vertical axis is the next. Preserves global spread but smears fine-grained clusters.
Scroll to zoom · drag to pan · hover to discover related work · click for details