More Than Words: Multimodal Interaction

A little while ago I released a package for dirichlet process mixture models in matlab. I'd like to use the blog to handle questions, as I did with the annealed particle filter. First of all, I'm able to get it to run by saying:


addpath('~/drg/lib/FullBNT-1.0.2/BNT/')
add_BNT_to_path
addpath(' ~/drg/lib/lightspeed/')
demodpmm


Or, for BNT 1.0.4


addpath('~/drg/lib/FullBNT-1.0.4');
addpath(genpathKPM('~/drg/lib/FullBNT-1.0.4'));
addpath('~/drg/lib/lightspeed/');
demodpmm


Please try this before emailing :)

Also, I was asked about the meaning of the variables:

On Wed, Jan 23, 2008 at 3:33 AM, Tian,Guangjian wrote:

"Would you please give me some reference for standard gaussian dirichlet process
mixture models?"

Jacob: I can recommend Rasmussen's "Infinite Gaussian Mixture Model" from NIPS 2000, Teh et al "Hierarchical Dirichlet Processes", and Erik Sudderth's PhD thesis, chapter 2.

Guangjian: "I find your matlab code for DP mixture models, I want to test the cluster
property of DP mixture models. But, when I use the code, I got some
errors(your example is ok), I want to fix the problem. But I don't know some
variables meanings in your program. Just like the following params."

params(1).alpha = T / 50; %1 / wishrnd(1,1);

J: concentration parameter -- see any tutorial on DPMMs, e.g. Rasmussen

params(1).kappa = .1; %T / 1000; %a pseudo count on the mean

J: Essentially, how much to trust the prior on the mean.

params(1).nu = 6; %a pseudo-count on the covariance

J: how much to trust the prior on the covariance

params(1).initmean = allmean;

J: prior on the mean of each cluster

params(1).initcov = allcov / 10;

J: prior on the covariance of each cluster

params(1).num_classes = 0;

J: system variable to keep track of the number of clusters generated

params(1).counts = 0;
params(1).sums = [];
params(1).cholSSE = [];

J: these are sufficient statistics for the cluster centers that can be incrementally updated. the mean of a cluster equals its sum divided by its count. The cholSSE variable is the cholesky decomposition of the covariance matrix.

params(1).classes = ones(T,1);

J: assignment of instances to clusters

% params(1).SSE = [];
params(1) = addNewClass(params(1));

J: initializes the first cluster.

params(1) = unhideObservations(params(1), 1, data);

J: assigns all observations to the first cluster
----

Good luck, and please post questions to the blog.
--Jacob