Hidden Markov Models with Nonparametric Emissions

We provide two learning algorithms for HMMs with nonparametric emissions: a naive nonparameteric EM and the spectral learning algorithm based on decompositions of continuous matrices (as described in the original paper). Continuous linear algebra is implemented using chebfun. KDE estimators are based on if-estimators.

Download

To get the code, clone the repository as follows:

$ git clone --recursive https://github.com/alshedivat/nphmm.git

Note: You need to clone recursively since nphmm depends on chebfun as a submodule.

Alternatively, the code can be donwloaded directly as a zip-file.

Installation

To use the library, add lib to the MATLAB path and setup chebfun as follows:

>> addpath(genpath('lib'))
>> chebfun_setup()

To ensure that the library is functional, you can run tests as follows:

>> runtests('tests')

Usage

The library provides two main classes, npHMM and npObsHMM, that correspond the nonparameteric HMM and its observable representation, respectively. Additionally, learnNPHMM trains a model from the provided data using either EM or spectral method. For usecases and examples check the scripts in examples/ and tests/.

Citation

@article{kax2016nphmm,
  title={Learning HMMs with Nonparametric Emissions via Spectral Decompositions of Continuous Matrices},
  author={Kandasamy, Kirthevasan and Al-Shedivat, Maruan and Xing, Eric P},
  journal={arXiv preprint arXiv:1609.06390},
  year={2016}
}

License

MIT (for details, please refer to LICENSE)