Description:

LIBIRWLS is an integrated library that incorporates a parallel implementation of the Iterative Re-Weighted Least Squares (IRWLS) procedure, an alternative to quadratic programming (QP), for training of Support Vector Machines (SVMs). Although there are several methods for SVM training, the number of parallel libraries is very reduced. In particular, this library contains solutions to solve either full or budgeted SVMs making use of shared memory parallelization techniques:

A parallel SVM training procedure based on the IRWLS algorithm.

A parallel budgeted SVMs solver based on the IRWLS algorithm.

For a detailed explanation of the algorithms take a look at the web page

Motivation

SVMs are a very popular machine learning technique because they can easily create non-linear solutions by transforming the input space onto a high dimensional one where a kernel function can compute the inner product of a pair vectors. Thanks to this ability, they offer a good compromise between complexity and performance in many applications.

SVMs have two main limitations. The first problem is related to their non-parametric nature. The complexity of the classifier is not limited and depends on the number of Support Vectors (SVs) after training. If the number of SVs is very large we may obtain a very slow classifier when processing new samples. The second problem is the run time associated to the training procedure that may be excessive for large datasets.

To face these problems, we can make use of parallel computing, thus reducing the run time of the training procedure or we can use budgeted approximations than can limit the complexity of the model in advance, which directly implies a faster classifier.

The above situation motivated us to develop "LIBIRWLS", an integrated library based on a parallel implementation of the IRWLS procedure to solve non-linear SVMs, either full and budgeted. This library is implemented in C, supports a wide range of platforms and also provides detailed information about its programming interface and dependencies.

Requirements:

This software is implemented in C and requires the following libraries:

OpenMP To parallelize the software.
ATLAS: Linear algebra package with standard routines that contains optimized BLAS and LAPACK implementations.

License: MIT License

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.