/scnn

Scalable Convex Neural Networks

Primary LanguagePythonMIT LicenseMIT

scnn: Scalable Convex Neural Networks

A library for global optimization of shallow neural networks. API documentation at ReadTheDocs.

Requirements

Python 3.8 or newer. Development dependencies are listed in dev_requirements.txt.

Setup

Install using pip:

python -m pip install pyscnn

Or, clone the repository and manually install:

git clone https://github.com/pilancilab/scnn.git
python -m pip install ./scnn

Contributions

Coming soon!

Citation

Please cite our paper if you use this package.

@article{DBLP:journals/corr/abs-2202-01331,
  author    = {Aaron Mishkin and
               Arda Sahiner and
               Mert Pilanci},
  title     = {Fast Convex Optimization for Two-Layer ReLU Networks: Equivalent Model
               Classes and Cone Decompositions},
  journal   = {CoRR},
  volume    = {abs/2202.01331},
  year      = {2022},
  url       = {https://arxiv.org/abs/2202.01331},
}

Looking for the code to replicate our experiments? See scnn_experiments.