- 26apr09 release

Cody has got Gibbs and collapsed Gibbs working for mixtures of
Gaussians. I have factored out the EM code from Mixtures. Various
other changes that I can't remember.


- 1apr09 release

I have changed from numbering the releases to naming them
by the release date.
This avoids arbitrary decisions about what constitutes a "major"
or "minor" release. 
There is a convention for such numbering, see:
http://www-subatech.in2p3.fr/~photons/subatech/soft/srt/html/c35283745.html.
However, I was not following it.
The pmtk convention is: if it is packaged as a zip file, it should be stable
and pass all tests; if you check it out from google code, it may be
buggy.

This new version contains code for inference in Gaussian junction
trees, Chow-Liu tree structure learning, preliminary code for Gibbs
sampling in GMMs, etc. Also, I have bundled MnistAll.mat and
data45.mat so you don't have to download them separately.


- 1.4.3 Released 17 March 09

Fixed various bugs. Now jtree returns logZ and can handle
hidden cts leaves.


- 1.4.2 Released 12 March 09

Now GmDist knows which nodes are discrete and cts,
and the node sizes (nstates). ConvertToTabularFactors interface
has changed.  lingaussHybridDemo updated.
familyTreeData and familyTreeGeneDataMystery
updated to use right noise model.


- 1.4.1 Released 7 March 09

Finished modifying interface to inference in joint distributions
(graphical models and MvnDist).
Now models are stateless, but engines are not.
The new interface is  as follows:

Unconditional marginals:
 
pQ  = marginal(model, queryNodes)  % p(Q)
pQs  = marginal(model, {Q1, Q2,...} ) % pQs{i} = p(Q{i})
 
Conditional marginals:
 
pQ  = marginal(model, queryNodes, visVars, visVals)  % p(Q|Visvars=visvals)
pQs  = marginal(model, {Q1,Q2,...}, visVars, visVals) % pQs{i} = p(Q{i}|V=v)

 
As before, different "engines" can be used to perform inference.
These can be specified as follows (same as before)
 
model.infEng = VarElimInfEng() % all hiddens are discrete
model.infEng = JtreeInfEng() % all hiddens are discrete
model.infEng = EnumInfEng() % all nodes are discrete
model.infEng = GaussInfEng() % all nodes are Gaussian
 
When calling marginal(model,...), the model passes the request
off to the inference engine, which deals with it. If you want to make
multiple marginal computations, you should do them in batch, using the cell
array syntax above, for efficiency.
 
 
 

- 1.4.0 Released 3 March 09

Flattened examples directory.

I am in the process of eliminating inference engine objects, and making the
relevant models call the underlying functions directly.
The engines have not yet been fully removed from the code,
but some of the functionality has been moved to GmDist,
and the demos in DgmDist now use the new conditional interface.
This release may be an inconsistent state; I have not run the test
suite yet. The reason for this change is that it simplifies the code a
lot; in particular, MvnDist and GmDist are much easier to understand.


- 1.3.3 Released 14 Feb 09

New testPMTK, which is faster than runDemos.
Gibbs sampling for Ising models.
Fixed bug in Gibbs sampling for discrete MRFs.


- 1.3.2 Released 4 Feb 09

Removed all .mex files, changed install procedure (loadPMTK and
compilePMTKmex), fixed various bugs, etc.


- 1.3.1 Released 1 Feb 09

Fixed svn problem, new version has util directory.
data45.mat removed to save space.

- 1.3 Released 29 Jan 09

VarElim now supports conditioning on cts nodes.
HMM casino demo modified. KNNdist.
Other stuff I can't remember.

- 1.2 Released 26 Jan 09

Now has more support for graphical models (eg variable elimination
inference engine, forwards-backwards infernece engine, fast factors).
Directory structure cleaned up. Etc.

- 1.1 Released 19 Dec 08

Separated Bayesian from MAP estimation by introducing compound
distributions. This version corresponds to the version presented at
the NIPS workshop.

- 1.0.3 Released 5 Dec 08.
Added HMMs with general class conditional densities.
Added DgmDist.
Got rid of VecDist, added ProductDist.
Many other changes.

- 1.0.2 Released 19 Nov 08. New ModelSelection class. All class names
start with Capital letter. Fit/inferParams merged. Many other changes.

- 1.0.1 Released 1 Nov 08.  Now all the demos are in the examples
directory, and are no longer static methods. Many other changes.

- 1.0.0 Released 26 Oct 08