Research map / Learning theory and algorithms / Prediction and decision losses
Bringing UMAP Closer to the Speed of Light with GPU Acceleration
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace CPU UMAP-learn with GPU-accelerated cuML UMAP | UMAP · MNIST, unsupervised | runtime: 52.575 ± 1.1677 -> 0.70781±0.0088 | Helped |
| Replace GPUMAP with cuML UMAP | UMAP · scRNA, unsupervised | trustworthiness: 94.35 -> 97.81 | Helped |
| Replace GPUMAP with cuML UMAP | UMAP · CIFAR-100, unsupervised | trustworthiness: 84.01 -> 83.42 | Hurt |
| Replace UMAP-learn with cuML UMAP | UMAP · Digits, unsupervised | trustworthiness: 98.79 -> 98.77 | No clear effect |
| Fuse adaptive smoothing operations instead of using separate intermediate-result kernels (ablation) | cuML UMAP · Adaptive smoothing operations | speedup: 12-15× | Helped |
| Accumulate source-vertex updates in registers (ablation) | cuML UMAP · Embedding optimization, n_components=2 | performance gain: 49% | Helped |
| Accumulate source-vertex updates in a shared-memory local cache (ablation) | cuML UMAP · Embedding optimization, n_components=2 to 16 | performance gain: 10% ... 56% | Helped |
| Use 64-bit accumulation and apply updates at each epoch end for exact reproducibility (ablation) | cuML UMAP · Embedding optimization, n_components=2 to 16 | runtime slowdown: 11× ... 20× | Hurt |
| Train on a 3% random sample, then embed the remaining data with distributed inference | Distributed UMAP · TASIC2018, approximately 23k cells | trustworthiness drop: less than a 1% drop | Hurt |
| Embed the remaining 97% over 16 workers | Distributed UMAP · TASIC2018, approximately 23k cells | trustworthiness drop: 0.05% | Hurt |
Closest papers by meaning
- FlashAttention-3: Fast and Accurate Attention with Asynchrony and Low-precision (NeurIPS 2024 · similarity 0.83)
- Tensorion: A Tensor-Aware Generalization of the Muon Optimizer (NeurIPS 2026 · similarity 0.82)
- StelLA: Subspace Learning in Low-rank Adaptation using Stiefel Manifold (NeurIPS 2025 · similarity 0.82)
- V-PETL Bench: A Unified Visual Parameter-Efficient Transfer Learning Benchmark (NeurIPS 2024 · similarity 0.81)
- DOGE-Train: Discrete Optimization on GPU with End-to-End Training (AAAI 2024 · similarity 0.81)
- Intrinsic Muon: Spectral Optimization on Riemannian Matrix Manifolds (NeurIPS 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.