Support vector machines for identifying organisms — a comparison with strongly partitioned radial basis function networks

Biological identification poses many problems which are not commonly met in other areas where artificial neural networks (ANNs) are applied. One such problem is that the number of possible classes is often unbounded and it is necessary to be able to deal with novel categories (taxa) either by rejecting them or retraining to allow for them. Where it is necessary to incorporate them, retraining a neural network may be time consuming and or impossible. One way of negating this problem is to use single-category networks that can be combined and configured as required. aper compares support vector machines (SVMs) with strongly partitioned traditional radial basis function (RBF) networks for the discrimination of single species of phytoplankton against a background of N other species. SVMs resulted in greater identification success than the unpartitioned, large single RBF networks (81 and 77%, respectively), though partitioned RBF ANNs performed relatively poorly (50% success) at discrimination of 61 marine phytoplankton species from flow cytometry data. Greatest success was achieved by combining the outputs of the individual networks by means of a ‘winner takes all’ strategy; with RBF ANNs identification success dramatically increased (from 16 to 50%), though there was only a modest increase with SVMs (77–81%). When SVMs trained on one data set were tested with data on cells grown under different light conditions, an overall successful identification was low (13%), but when SVMs were trained on a combined data set identification was high (>70%). Clearly, it is essential to cover the whole spectrum of biological variation in the training data set.