REVE Embedding Atlas

Interactive browser-based explorer for the REVE pretraining dataset with 1M+ projected EEG/MEG windows using DuckDB-WASM

The REVE Embedding Atlas is an interactive, serverless web application that allows researchers to visually explore and analyze projected EEG and MEG windows from the REVE pretraining dataset (25,000 subjects across 92 datasets) directly in the browser.
Interactive 2D embedding atlas powered by Apple Embedding Atlas and DuckDB-WASM.

Highlights

  • Interactive Embedding Exploration: Visualizes projected 2D embeddings of over 1,000,000 EEG/MEG windows using Apple Embedding Atlas, color-coded across diverse datasets (TUH, ICARE, OpenNeuro, etc.).
  • Completely Serverless (DuckDB-WASM): Queries public Parquet tables hosted on Hugging Face directly inside the browser using DuckDB-WASM and HTTP byte-range requests — zero backend server required.
  • Subset & Diversity Selection: Explores Determinantal Point Processes (DPP) and ranking selections across candidate pools.
  • Live Cross-Filtering: Real-time SQL predicate filtering, reconstruction loss histograms, and dataset breakdowns.

References