loading page

Synthetic Neural Network: Weight Divergence Optimizer
  • Alwaleed Mohamed
Alwaleed Mohamed
Author Profile

Abstract

Any endeavors to explain the network behavior, always falls in the network optimizer explanation trap and gets into the complication of algorithm math operation loop without reaching the final decision of how it really works. Weight divergence optimizer, which is a subcomponent of Synthetic Neural Network explains the inner works of neural network in a pure logical operation using only the dot product multiplication to illustrate this behavior and proposed a formula to calculate bias and weights. The new method uses the network divergence theory instead of trial-and-error method by calculating the maximum variation value between weight and its interior class patterns compared with the minimum variation value of the same weight with exterior patterns of other classes, the difference between the two values used in the formula. Synthetic Neural Network deals with many challenges starting from how Neural Network works and get its result, extending to reducing the size of training dataset, huge memory consumption, extensive processor calculation and creditability. The network output shows a promising result in testing and validation stage, it grants an accuracy of 90% to 75% with two and nine classes when using USPS dataset. Also, it produces accuracy of 95% to71% when using RMNIST expanded with the same classes. In audio patterns with MFCC feature extractions it gives accuracy of 73% with 10 classes. All the training operation performed between 0.5 to 15.5 seconds. The proposed method uses the lowest number of neurons that ever used with two levels of Neural Network, that helps to reduce the network size, discards redundant patterns, removes corrupted inputs and makes the network to converge in a few seconds with the lowest number of input dataset used