Distributed SVM learning and support vector reduction

A Support Vector Machine (SVM) is a machine learning classification method with robust learning and minimal overtraining concerns. In this paper experimental data involving a nanopore detector is examined, where 150component feature data is gathered on individual molecules. SVM training data can be produced by the nanopore detector in prodigious amounts, to arrive at a set of 150-component feature vectors that number from 10,000 to 100,000 for the experiments described, depending on the number of molecular classes being examined for a particular application. Training sets of 10,000 or more, however, can’t be managed with a single PC-based resource. For this reason most SVM implementations must contend with some kind of chunking process to learn parts of the data at a time. In this paper, two sets of binary SVM training results are examined. These results show that chunk aliasing and outlier accumulation may pose problems for distributed SVM learning. The results also present new methods and how they offer a stable learning solution to these problems at minimal cost. One of the methods extends the learning process with modified alpha-selection heuristics that enable a support-vector reduction phase. The distributed SVM learning described here was implemented using Java RMI, and was developed to run on a network of multi-core computers. 92 Stephen Winters-Hilt