Superposition Replication Study

This repository contains implementations and experiments exploring superposition phenomena in different neural network architectures. Superposition is a phenomenon where neural networks learn to encode multiple features in the same set of weights or neurons, effectively compressing information through overlapping representations.

Project Structure

├── images/                               # Visualization outputs and diagrams
├── runs/                                 # Training logs and experiment results
├── scripts/                              # Utility scripts and helpers
├── anthropic_toy_models.ipynb            # Jupyter notebook with toy model implementations
├── environment.yaml                      # Conda environment specification
├── intro_transformer_superposition.py     # Basic transformer superposition examples
├── intro_translation_superposition.py     # Translation model superposition examples
├── toy_models_config.yaml                # Configuration for toy model experiments
├── toy_models_reproduction.py            # Scripts to reproduce toy model results
├── transformer_superposition.py           # Advanced transformer implementations
├── translation_superposition.py          # Advanced translation model experiments
└── LICENSE                               # Project license

Getting Started

Prerequisites

To run the experiments, you'll need Python 3.7+ and conda installed. Set up the environment using:

conda env create -f environment.yaml
conda activate superposition

Running Experiments

The repository includes several implementations:

1. Basic Examples

   python3 intro_transformer_superposition.py
   python3 intro_translation_superposition.py

   Demonstrates superposition patterns in transformer and translation architectures.

2. Toy Model Reproduction

   python3 toy_models_reproduction.py

   Reproduces results from toy model experiments using configurations in toy_models_config.yaml.

3. Advanced Experiments

   python transformer_superposition.py
   python translation_superposition.py

   Contains more sophisticated experiments and analysis tools.

Key Features

• Implementation of superposition detection methods
• Interactive toy models for understanding basic concepts
• Visualization tools for analyzing learned representations
• Comparative analysis across different model architectures
• Experiments with various training configurations
• Tools for measuring and quantifying superposition effects

Visualization and Analysis

Results and visualizations are stored in the images/ directory. Training logs and metrics can be found in the runs/ directory.

Contributing

Contributions are welcome! Please feel free to submit a Pull Request. For major changes, please open an issue first to discuss what you would like to change.

License

This project is licensed under the terms included in the LICENSE file.

Citation

If you use this code in your research, please cite:

@software{superposition_replication,
  title = {Model Superposition Replication Study},
  year = {2024},
  url = {https://github.com/mmjerge/superposition_replication}
}