nlprinceton / sparse_recovery

noiseless/nonnegative sparse recovery and feature retrieval via compressed sensing

sparse_recovery

This module provides solvers and utility functions for several problems in sparse recovery/compressed sensing. The file solvers.py, which depends only on NumPy/SciPy and can be used with Python 2 or 3, provides methods to compute the following:

• Basis Pursuit (BP), including Nonnegative BP (BP+); the BP solver is ported from the l1-MAGIC MATLAB package [1]; the BP+ solver is based on the same primal-dual interior point method [2].
• Orthogonal Matching Pursuit (OMP), including Nonnegative OMP (OMP+).
• Supporting Hyperplane Property (SHP), a property that guarantees recovery of a signal x from linear measurements Ax via BP+ [3].

The feature retrieval files (retrieval.py, word_embeddings.py) additionally require scikit-learn and text_embedding. These files are used by scripts in the directory scripts-AKSV2018 to recreate the results in [3].

If you find this code useful please cite the following:

@inproceedings{arora2018sensing,
  title={A Compressed Sensing View of Unsupervised Text Embeddings, Bag-of-n-Grams, and LSTMs},
  author={Arora, Sanjeev and Khodak, Mikhail and Saunshi, Nikunj and Vodrahalli, Kiran},
  booktitle={Proceedings of the 6th International Conference on Learning Representations (ICLR)},
  year={2018}
}

References:

[1] Candès & Romberg, "l1-MAGIC: Recovery of Sparse Signals via Convex Programming," Technical Report, 2005.

[2] Boyd & Vandenberghe, "Chapter 11: Interior-point Methods," Convex Optimization, 2004.

[3] Arora et al., "A Compressed Sensing View of Unsupervised Text Embedding, Bag-of-n-Grams, and LSTMS," ICLR, 2018.