A Comparative Study of Dendrite Neural Networks for Pattern Classification

Abstract

Dendrite neurons are an alternative for classification tasks, providing competitive results when compared to typical classification methods. Dendrite networks allow each dendrite to build a close boundary to assign each incoming pattern to its respective class. Hyperboxes, hyperellipsoids and hyperspheres are novel ways for dendrite computing. In this research we test these models and some hybrid variances trained by stochastic gradient descent. Results show that hyperellipsoid work well as classifiers with low-dimensional tasks, while hyperspheres score better than the others in the case of image processing. However, when hybridizing, hyperboxes show poor results but hyperellipsoid and hyperspheres obtain even better results than two layer prerceptrons for many datasets.