The Dendrite Morphological Neurons Improving Classification Performance

Pattern categorization neural models are dendritic morphological neurons (DMNs), which are geometric shapes that enclose patterns of the same class. This research examines how box, ellipse, as well as sphere dendrite shapes affect pattern categorization. To further refine the decision-making boundaries generated by conventional DMNs, we propose using a softmax layer attached to the DMN's output to calculate posterior likelihoods using weighted dendritic responses. Automatically, an incremental-decremental tuning method is used to fine-tune the total number of dendrites for each category. Classification effectiveness is measured over 49 real-world and 9 synthetic datasets. Meanwhile, the precision and complexity of the model are evaluated across 12 distinct DMN variants. The DMN's potential is enhanced by its spherical dendrites, its smooth activation mechanisms, and its easily accessible softmax layer. It is the most precise, has the simplest geometric shape, is unaffected by parameters with little variation, and has a simpler structure thanks to the smooth maximum function. When compared to other popular machine learning methods, this DMN setup achieved higher accuracy rates than the support vector machines. multilayer perceptron, radial base function system, k-nearest neighbors, and random forest. As a result, the DMN is an effective pattern classification approach for problems in the real world.