CUED Publications database

Markov beta processes for time evolving dictionary learning

Shah, A and Ghahramani, Z (2016) Markov beta processes for time evolving dictionary learning. In: UNSPECIFIED pp. 676-685..

Full text not available from this repository.


We develop Markov beta processes (MBP) as a model suitable for data which can be represented by a sparse set of latent features which evolve over time. Most time evolving nonparametric latent feature models in the literature vary feature usage, but maintain a constant set of features over time. We show that being able to model features which themselves evolve over time results in the MBP outperforming other beta process based models. Our construction utilizes Poisson process operations, which leave each transformed beta process marginally beta process distributed. This allows one to analytically marginalize out latent beta processes, exploiting conjugacy when we couple them with Bernoulli processes, leading to a surprisingly elegant Gibbs MCMC scheme considering the expressiveness of the prior. We apply the model to the task of denoising and interpolating noisy image sequences and in predicting time evolving gene expression data, demonstrating superior performance to other beta process based methods.

Item Type: Conference or Workshop Item (UNSPECIFIED)
Divisions: Div F > Computational and Biological Learning
Depositing User: Cron Job
Date Deposited: 17 Jul 2017 19:19
Last Modified: 19 Jul 2018 07:26