Deep Unfolding Bayesian Non-Negative Matrix Factorization

Bayesian non-negative matrix factorization is a method used across fields including genomics, audio and signal processing, and neuroscience. The complexity of the posterior of Bayesian NMF requires MCMC methods, such as a Gibbs sampler, or variational inference. We propose a faster solution through deep algorithm unrolling. By designing a neural network where each layer mimics a single iterative update, we are able to improve speed without sacrificing model performance.

---

Advised by Demba Ba, PhD
CRISP Lab, Harvard School of Engineering and Applied Sciences.