Timur

Would be nice to have some more involved methods for the framework, such as Bayesian PCA (see http://research.microsoft.com/apps/pubs/default.aspx?id=67157 )

currently, only constant parents for both gamma and Wishart distributions are supported. This does not allow hierarchical variance modelig e.g. like in http://arxiv.org/abs/1207.1380 It is possible to tackle this problem...

speed up

1

training a model with 32 mixtures currently takes around 1.5 seconds. since networks are entirely independent, it should be possible to multithread (c++11 ThreadPool implementation - https://github.com/en4bz/ThreadPool) Also, when fitted,...

Tests for basic nodes are needed, as well as for basic examples like gaussian mixtures. Both for C++ and later for matlab bindings

Think of how to automatically generate inference algorithm. Either dynamically using the structure of the network or using code generation.

Make a hard-coded VMP-based implementation of algonquin.

object tracking example

1

create a demo application - object tracking using the toolbox. something like this: http://www.mathworks.nl/products/matlab-coder/examples.html?file=/products/demos/shipping/coder/coderdemo_kalman_filter.html

dynamic networks support

1

Currently one has to manually propagate messages from step to step for a dynamic network. Maybe extending `Network` with something that allows to specify connections between different (time)steps could work.

Add checks for the number / type / dimensionality consistency of input arguments for `mexFunction`

just make them better. look at matlab publish?