e-means paper trick: IVF approximate assignment cuts clustering cost 34x at 10B vectors
antoine_chaffin · x · 2026-11-10
Researcher antoinechaffin highlights a method from the e-means paper: index-accelerated approximate assignment. With millions of centroids, exact distance computation becomes prohibitive, so centroids are indexed with IVF (bucketed by meta-centroids) and treated as approximate nearest-neighbor search.
GPU hours to reach Lloyd's test loss on DINO10B subsets: no gain at 100M vectors, 4.3x cheaper at 1B, and 34x cheaper at 10B — training 16.7M centroids in under one GPU-day. He also submitted a PR to TACHIOM supporting the same technique (cpu, HNSW, non-batched) as an alternative to TAC clustering.
More from Infra
- Fireworks AI hacked: customer data and models compromised, customers flee — annatonger · 2026-11-10
- AI's bottleneck shifts from chips to power: model firms spend billions on turbines — TansuYegen · 2026-11-10
- How CrossnoKaye grounds physical AI agents in closed-loop industrial control with Bedrock AgentCore — AWS ML Blog · 2026-11-10
- Gary Marcus slams Nvidia after report says Commerce inquiry was dropped — GaryMarcus · 2026-11-10
- Stratoscope: monitor local LLM token speed and cache hits live in your menu bar — LisaMarieJosefine · 2026-11-10
- Claude Fable 5.2 and Mythos 5.2 spotted on AWS Bedrock, priced 20% below 5.1 — testingcatalog · 2026-11-10