Overcomplete is a compact research library in Pytorch designed to study (Overcomplete)-Dictionary learning methods to extract concepts from large Vision models. In addition, this repository also introduces various visualization methods, attribution and metrics. However, Overcomplete emphasizes experimentation.
Overcomplete requires Python 3.8 or newer and several dependencies, including Numpy. It supports both only Torch. Installation is straightforward with Pypi:
pip install overcompleteWith Overcomplete installed, you can dive into an optimisation based dictionary learning method to extract visual features or use the latest SAEs variant. The API is designed to be intuitive, requiring only a few hyperparameters to get started.
Example usage:
import torch
from torch.utils.data import DataLoader, TensorDataset
from overcomplete.sae import TopKSAE, train_sae
Activations = torch.randn(N, d)
sae = TopKSAE(d, nb_concepts=16_000, top_k=10, device='cuda')
dataloader = DataLoader(TensorDataset(Activations), batch_size=1024)
optimizer = torch.optim.Adam(sae.parameters(), lr=5e-4)
def criterion(x, x_hat, pre_codes, codes, dictionary):
mse = (x - x_hat).square().mean()
return mse
logs = train_sae(sae, dataloader, criterion, optimizer,
nb_epochs=20, device='cuda')- Getting started:
- TopK, BatchTopK, JumpReLU, Vanilla SAE:
- Stable Dictionary with Archetypal-SAE:
- Advanced metrics to study the solution of SAE:
- The visualization module:
- NMF, ConvexNMF and Semi-NMF:
- Modern Feature visualization to visualize concepts: Coming soon
This work has been made possible in part by the generous support provided by the Kempner Institute at Harvard University. The institute, established through a gift from the Chan Zuckerberg Initiative Foundation, is dedicated to advancing research in natural and artificial intelligence. The resources and commitment of the Kempner Institute have been instrumental in the development and completion of this project.
For a complete LLM implementation of the SAE, we strongly recommend exploring the following resources. The Sparsify library by EleutherAI provides a comprehensive toolset for implementing the SAE. The original TopK implementation is available through OpenAI's Sparse Autoencoder. Additionally, SAE Lens is an excellent resource, especially if you are interested in using the SAE-vis. Finally, closer to our work, for those interested in the vision domain, ViT-Prisma is an excellent mechanistic interpretability library for Vision Transformers (ViTs) that supports activation caching and SAE training.
The Overcomplete framework has contributed to the following research publications.
@article{fel2025archetypal,
title = {Archetypal SAE: Adaptive and Stable Dictionary Learning for Concept Extraction in Large Vision Models},
author = {Fel, Thomas and Lubana, Ekdeep Singh and Prince, Jacob S. and Kowal, Matthew and Boutin, Victor and Papadimitriou, Isabel and Wang, Binxu and Wattenberg, Martin and Ba, Demba and Konkle, Talia},
journal = {arXiv preprint arXiv:2502.12892},
year = {2025},
url = {https://arxiv.org/abs/2502.12892}
}
@article{thasarathan2025universal,
title = {Universal Sparse Autoencoders: Interpretable Cross-Model Concept Alignment},
author = {Thasarathan, Harrish and Forsyth, Julian and Fel, Thomas and Kowal, Matthew and Derpanis, Konstantinos},
journal = {arXiv preprint arXiv:2502.03714},
year = {2025},
url = {https://arxiv.org/abs/2502.03714}
}
@article{hindupur2025projecting,
title = {Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry},
author = {Hindupur, Sai Sumedh R. and Lubana, Ekdeep Singh and Fel, Thomas and Ba, Demba},
journal = {arXiv preprint arXiv:2503.01822},
year = {2025},
url = {https://arxiv.org/abs/2503.01822}
}
- Thomas Fel - tfel@g.harvard.edu, Kempner Research Fellow, Harvard University.
