Background and description
Artificial neural networks (ANNs) are a branch of machine learning models that are built using principles of neuronal organization, based in a collection of connected units or nodes (artificial neurons), which loosely model the neurons in a biological brain. ANNs learn (are trained) by processing examples, each of which contains a known “input” and “result”, forming probability-weighted associations between the two, which are stored within the data structure of the net itself. Currently, almost all training techniques of artificial neural networks use the method of initialization of the weights of neurons relying on random statistical distributions, such as Glorot Uniform, Normal Glorot, Random Uniform or Random Normal distributions, among many others. These random distributions have proven to be excellent choices for initializing the weights of an ANN's neurons before starting training.
The present invention refers to a new deterministic method, with perfect sequences (ANN PS), that greatly improves and accelerates the training speed of an ANN by a factor greater than x10 compared to other known stochastic method of initiating the weights of neurons. ANN PS method outperforms all the other random methods mentioned earlier. This new method consists of initializing the weights of each neuron of an ANN with specific values and the increment of training speed is as higher as the number of hidden layers in the ANN.
