new tech report and new version of SVMTorch

[Samy Bengio]

We would like to announce the following:

Regarding SVMTorch, our implementation of Support Vector Machines for
Large-Scale Regression and Classification Problems, that was previously
announced on the same list,

(a) A new version of SVMTorch is now available on the web at the usual place: 

            http://www.idiap.ch/learning/SVMTorch.html

    The main additions compared with the previous version are the following:
    + A Multiclass mode (one class against the others).
    + An input/output sparse mode (it runs faster on sparse data).

(b) A new technical report is also available reporting a convergence proof
    for our regression method. It is available at 

            ftp://ftp.idiap.ch/pub/reports/2000/rr00-17.ps.gz

    The abstract is as follows:

    Recently, many researchers have proposed decomposition algorithms for
    SVM regression problems (see for instance [11, 3, 6, 10]). In a previous 
    paper [1], we also proposed such an algorithm, named SVMTorch. In this 
    paper, we show that while there is actually no convergence proof for any 
    other decomposition algorithm for SVM regression problems to our knowledge,
    such a proof does exist for SVMTorch for the particular case where no 
    shrinking is used and the size of the working set is equal to 2, which is 
    the size that gave the fastest results on most experiments we have done. 
    This convergence proof is in fact mainly based on the convergence proof 
    given by Keerthi and Gilbert [4] for their SVM classification algorithm.


-----
Samy Bengio  
Research Director. Machine Learning Group Leader.
IDIAP, CP 592, rue du Simplon 4, 1920 Martigny, Switzerland.
tel: +41 27 721 77 39, fax: +41 27 721 77 12.
mailto:bengio at idiap.ch, http://www.idiap.ch/~bengio