Meta FAIR's e-means reaches Lloyd's 25-epoch loss in 1.4 to 2.6 epochs, shipped in FAISS 1.16.0
Meta FAIR's e-means is a mini-batch k-means replacement shipped in FAISS 1.16.0. According to MarkTechPost, it reaches Lloyd's 25-epoch loss in 1.4 to 2.6 epochs, with 10-45x GPU speedups and no extra hyperparameters.
Meta FAIR's e-means, a mini-batch k-means replacement, is now available in FAISS 1.16.0. MarkTechPost reports that it reaches the loss Lloyd's algorithm attains after 25 epochs in 1.4 to 2.6 epochs.
Key details
- What it is: e-means is a replacement for mini-batch k-means, developed by Meta FAIR.
- Where it ships: FAISS 1.16.0.
- Convergence: it reaches Lloyd's 25-epoch loss in 1.4 to 2.6 epochs.
- Speed: 10-45x GPU speedups.
- Tuning: no extra hyperparameters.
The figures above are as reported by MarkTechPost in its coverage of the release. Details on benchmark datasets, hardware, and baseline configurations were not part of the summary.
What this means
k-means clustering is a standard step in vector search and indexing workflows, and FAISS is the library where e-means now ships. Reaching Lloyd's 25-epoch loss in 1.4 to 2.6 epochs means substantially fewer passes over the data for the same clustering loss, according to the reported figures. The absence of extra hyperparameters means that adopting e-means in place of mini-batch k-means does not add new settings to tune. Teams that train clusterings with FAISS can evaluate the algorithm by upgrading to version 1.16.0.
Related Articles
Amazon releases Mitra-v2, an open tabular foundation model that leads the single-model field on TabArena
Amazon has released Mitra-v2, a tabular foundation model for classification and regression that, according to Amazon, leads the single-model field on the TabArena benchmark of 51 datasets. Weights and code are available under Apache-2.0 on Hugging Face.
PSI used early access to Anthropic's Fable 5.2 to resolve the eighteen-year-old Aaronson-Ambainis conjecture
Physical Superintelligence (PSI) used early access to Anthropic's Fable 5.2 model to solve the Aaronson-Ambainis conjecture, an eighteen-year-old quantum computing problem. The startup earns revenue from data center physics simulation and wants to turn theoretical physics results into technologies.
Google DeepMind's AlphaProof Code formally verifies key Signal protocol guarantees in four weeks
Google DeepMind used a new agentic AI framework, AlphaProof Code, to mathematically prove a number of the Signal protocol's key cryptographic guarantees. The proof-of-concept took four weeks and found no new vulnerabilities. DeepMind is releasing the specification and proof, and WIRED reports the analysis did not cover the entire Signal mobile app.
ModelBest's LES runs 35B MoE MiniCPM-V 4.7 at over 20 tokens/s on iPhone 18 Pro Max, company says
ModelBest published an engineering write-up on Lookahead Expert Scheduling (LES), the inference system behind MiniCPM-V 4.7 35B-A3B. The company reports over 20 tokens per second on an iPhone 18 Pro Max, and has open-sourced the code in its llama.cpp-omni repository.
Comments
Loading...