SAE Semantic Explorer

A visualization tool for exploring SAE (Sparse Autoencoder) features using UMAP and ball mapper. Paper: Visual Exploration of Feature Relationships in Sparse Autoencoders with Curated Concepts — published at the Mechanistic Interpretability Workshop at NeurIPS 2025.

🎬 Demo Video

Data

1. Download the data from Google Drive
2. Unzip the downloaded file.
3. Move the extracted data folder into the backend directory so that the path is backend/data/.

Structure

• backend/ - Flask API server
• frontend/ - React web interface

Installation

Backend Setup

1. Navigate to the backend directory:

   cd backend

2. Create a virtual environment (recommended):

   python -m venv venv
   source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install Python dependencies:

   pip install -r requirements.txt

Frontend Setup

Tested environment:

Node.js version 18.20.8 (node -v)
npm version 10.8.2 (npm -v)

1. Navigate to the frontend directory:

   cd frontend

2. Install Node.js dependencies:

   npm install

Quick Start

1. Start the backend:

   cd backend
   python app.py

2. Start the frontend:

   cd frontend
   npm start

📚 Cite Our Work

@inproceedings{yan2025visual,
  title={Visual Exploration of Feature Relationships in Sparse Autoencoders with Curated Concepts},
  author={Yan, Xinyuan and Liu, Shusen and Thopalli, Kowshik and Phillips, Bei Wang},
  booktitle={Mechanistic Interpretability Workshop at NeurIPS 2025}
}