FIT Božetěchova 2, A112, 10:00 3.4.2009

The presentation will introduce Niko's new MATLAB toolkit. The toolkit has two main capabilities:

1. It has a powerful and efficient general purpose optimization algorithm for large-scale unconstrained optimization.

2. It has a framework to enable the user to synthesize (i) new pattern recognition systems and (ii) customizable objective functions, which together form new discriminative training algorithms. This is made possible by the capability of the toolkit to dramatically reduce the complexity of writing code to compute the first and second order partial derivatives which are necessary for the above optimizer to work.

While (2) is necessary for (1), (2) may be used independently for other purposes as well. For example, it may be used to compute the derivatives for other optimization (and discriminative training algorithms).

As part of (2), the toolkit has a library of elementary functions, for which the derivative code has already been implemented. These elementary functions can be combined together into more complex systems and objective functions. It is my hope that other users of the toolkit will help to write more such function implementations as they need them for their own purposes and to contribute them to the library for others to use.

Finally, the toolkit comes with a complete multiclass linear logistic regression implementation, which is suitable for large-scale training tasks. We used it to train a language recognizer, with a training set of the order of 104 trials and input feature dimension of the order of 105, in about 10 minutes.

Speakers