Multi-task learning framework (see Jayaram et al. 2016) written in MATLAB. Python version is in pyMTL
Switch branches/tags
Nothing to show
Clone or download

README.md

Transfer Learning in Brain-Computer Interfaces

Latest Version Downloads

Object-oriented multi-task learning framework for Bayesian hierarchial models written in MATLAB. Implements the regression based approaches from Jayaram et al. [1] and the logistic approach by Fiebig et al. [2], but should be easily extendable.

Installation

The folder simply needs to be put into the MATLAB path

Usage

The file testscript.m has sample data and runs through classification with the various approaches. Generally, however, the order of things is as follows:

  1. Instantiate the model with appropriate size parameters and switches
model = MT_linear([flags]);
moel = MT_FD_model({'linear','logistic', [flags])

Note that there are two possible models (for details see [1]): the linear and bilinear. The linear model requires datasets of the form (features x labels) while the bilinear model (accessible through MT_FD_model) requires datsets of the form (electrodes x features x labels).

  1. Train the prior (and optionally classify using the prior mean)
model.fit_prior(X_cell, y_cell)
# prior mean classification
y_hat = model.prior_predict(new_X)
  1. Train the subject specific model. The output struct includes a classification function as well as various explanatory parameters.
updated = model.fit_new_task(Xtrain, ytrain)
y_hat = updated.predict(new_X)

For more help information and information on flags, please check the documentation for the functions MT_baseclass and MT_linear

Functionality

Base class

MT_baseclass implements a class that can be inherited which sketches out the general form of the two algorithms: Dataset specific models are updated in parallel (when possible) and then a distribution is generated from the data-specific models, repeated in an alternating fashion. So, your basic E-M approach. Any class can inherit from this in order to create more algorithms of this nature

Dataset-specific models

MT_linear is the basic approach that considers the multi-task problem in a linear regression setting [1]. To create new methods, the easiest way is to inherit from the class and simply change the method that fits the classifier given a prior and a dataset (see MT_logistic for an example).

Additions

There is an enormous space of possibilties for how this framework can be extended and improved. If you are interested in adding a method please let me know (vjayaram@tue.mpg.de) and see the included CONTRIBUTE.md file for some help on how your method could be fit into this framework.

Feedback

Please feel free (indeed, urged) to let me know through the issues feature whether something is not working and I will be happy to fix them as soon as I can. If preferrable, feel free also to send mail to vjayaram@tue.mpg.de

Python

For the python Python version please check out our related page.

Citations:

[1] Jayaram, Vinay, et al. "Transfer learning in brain-computer interfaces." IEEE Computational Intelligence Magazine 11.1 (2016): 20-31.

[2] Karl-Heins Fiebig, Vinay Jayaram, Jan Peters, and Moritz Grosse-Wentrup. Multi- task logistic regression in brain-computer interfaces. In Proceedings of the 2016 IEEE International Conference on Systems, Man, and Cybernetics, 2016.