Skip to content

README.md

Latest commit

 

History

History
82 lines (59 loc) · 2 KB

README.md

File metadata and controls

82 lines (59 loc) · 2 KB

COMET: Conditionally-Overlapping Mixture of Experts

Paper: More Experts Than Galaxies: Conditionally-Overlapping Experts With Biologically-Inspired Fixed Routing

Overview

About this repository

This repository contains the official code for reproducing results from the paper:

"More Experts Than Galaxies: Conditionally-Overlapping Experts With Biologically-Inspired Fixed Routing"
by Sagi Shaier, Francisco Pereira, Katharina von der Wense, Lawrence E. Hunter, and Matt Jones.

The code enables training and evaluating sparse modular neural networks based on our proposed COMET architecture on benchmark datasets.

Prerequisites

  • Python 3.11
  • Conda (for environment management)

Installation

1. Clone the Repository

git clone https://github.com/Shaier/COMET.git
cd COMET

2. Set Up the Conda Environment

Create and activate a Conda environment with Python 3.11:

conda create -n COMET python=3.11
conda activate COMET

3. Install Dependencies

Install the required Python packages using pip:

pip install -r requirements.txt

Usage

Run Experiments

To reproduce standard model experiments (4-layer MLPs on CIFAR-10, illustrating the effect of capacity and sparsity — Figures 4 and 9 in the paper):

python run_models.py

Plot Results

To visualize experimental results:

python plot_results.py

Citation

If you use this codebase or refer to our paper, please cite:

@misc{shaier2025expertsgalaxiesconditionallyoverlappingexperts,
      title={More Experts Than Galaxies: Conditionally-overlapping Experts With Biologically-Inspired Fixed Routing}, 
      author={Sagi Shaier and Francisco Pereira and Katharina von der Wense and Lawrence E Hunter and Matt Jones},
      year={2025},
      eprint={2410.08003},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2410.08003}, 
}