GPUMAP: GPU-Accelerated Progressive Uniform Manifold Approximation and Projection

Abstract

We present GPU-accelerated Progressive Uniform Manifold Approximation and Projection (GPUMAP), a progressive implementation of UMAP for interactive large-scale dimensionality reduction. GPUMAP decomposes UMAP into schedulable operations, and returns intermediate projections under operation-count or time-budget constraints. We examine how these intermediate projections evolve toward final reference projections under different scheduling strategies and demonstrate progressive projection on a 5M-point large-scale dataset. Our results show that progressive execution can make large-scale DR more observable before full computation is complete.