Dendrite ellipsoidal neurons based on k-means optimization

Dendrite morphological neurons are a type of artificial neural network that can be used to solve classification problems. The major difference with respect to classical perceptrons is that morphological neurons create hyperboxes to separate patterns from different classes, while perceptrons use hyperplanes. In this paper, we introduce an improved version of dendrite morphological neural networks, which we have called dendrite ellipsoidal neuron that employs hyperellipsoids instead of hyperboxes. This ellipsoidal neuron is presented with a new training algorithm, to set the covariance matrix and the centroid of each hyperellipsoid based on k-means++, by applying hill climbing to search for an optimum number of hyperellipsoids. The main advantage of this approach is that dendrite ellipsoidal neuron creates smoother decision boundaries. The proposed neural model was tested on synthetic and real datasets from the UCI machine learning repository (in a paired t-test) achieving an average accuracy of 80.7%, while multi-layer perceptrons gave 78.4%, support vector machines obtained 74.2%, and radial basis networks 72.7%. Lastly, to test the proposed method performance in solving real practical problems, our model was used to detect lane lines on an urban highway, for classifying figures with a Nao robot and for traffic detection.