ibUMAP paper: coherent field evaluation speeds up UMAP up to 5.79x with better stability
leland_mcinnes · x · 2026-10-05
An arXiv paper (under review at ICLR 2027) proposes ibUMAP, replacing UMAP's stochastic negative sampling with a coherent field-based approach: attraction and repulsion are evaluated from a shared embedding snapshot and applied synchronously, with a degree-weighted repulsive field represented by three scalar moments and computed via an interpolation-based FFT scheme.
Key results:
- Median speedups of 3.29x unseeded and 5.79x seeded vs umap-learn on CPU; 1.44x over cuML on million-scale GPU runs
- Much better run-to-run stability
- Synchrony and kernel capping shift the local-global fidelity trade-off; FFT evaluation yields only small average quality changes
- Code released on GitHub
More from Research
- Single-layer transformer's MLP block invents a brand-new word — Sauers_ · 2026-10-05
- Ben Goertzel: What AGI, RSI and Superintelligence Originally Meant and How They Got Confused — bengoertzel · 2026-10-05
- Evolutionary LISP kernel uses LLMs to self-modify apps with zero-downtime hot reload — max_paperclips · 2026-10-05
- Neuroscientists say AI is accelerating brain science: a field report from CoSyNe 2026 — HistedLab · 2026-10-05
- Schmidhuber joins Sakana AI as chief scientific advisor, free Tokyo symposium Oct 26 — hardmaru · 2026-10-05
- Stanford kundajelab releases MotifCompendium, GPU-accelerated motif analysis toolkit — anshulkundaje · 2026-10-05