An Alternative Way of Determining Biases and Weights for the Training of Neural Networks

The determination of biases and weights in neural networks is a fundamental aspect of their performance, traditionally employing methods like steepest descent and stochastic gradient descent. While these supervised training approaches have proven effective, this technical note confidently presents a groundbreaking alternative that eliminates randomness in calculations altogether. This innovative method calculates biases and weights as precise solutions to a system of equations. This approach not only reduces computational demands but also enhances energy efficiency significantly. By strategically incorporating target values during training, we can expand the number of target values within acceptable variance limits, enabling the formation of square matrices. This ensures that input and output nodes remain perfectly balanced through the generation of fictive data, particularly for output nodes. These fuzzy sets guarantee that the variance of neuron target values stays within permissible limits, effectively minimizing error. The generated data is intentionally minimal and can also be produced using random processes, facilitating effective learning for the neural networks. Unlike conventional techniques, the values of biases and weights are determined directly, leading to a process that is both faster and less energy-intensive. Our primary objective is to establish an efficient foundation for the training data of the neural network. Moreover, these calculated values serve as robust initial parameters for integration with other determination methods, including stochastic gradient descent and steepest descent. This presentation showcases a powerful new algorithm, poised to significantly enhance the efficiency and effectiveness of neural network training.