A novel data-driven statistical learning algorithm for neural networks applied to language translation

Authors:

Abstract
This article presents a novel statistical data-driven learning algorithm for neural network training, designed for low-resource data scenarios. The proposed method employs statistical descriptors, including the sample mean, standard deviation, z-score, and normal distribution, to estimate the probability of prediction error and guide the adaptive updating of weights and biases. By incorporating error normalization and probabilistic mappings, the algorithm provides a statistically guided learning mechanism that accounts for signal variability while promoting stable convergence. Contrasting conventional approaches that rely primarily on gradient information, the proposed framework uses probabilistic information derived from the distribution of error to regulate parameter adaptation. The methodology is validated using Nahuatl–Spanish lexical translation task implemented with neural networks. • Proposes a statistical–probabilistic learning algorithm that uses error distributions instead of relying only on gradients.• The method is intended for low-resource scenarios and shows stable convergence under the evaluated conditions.• Validated on a Nahuatl–Spanish lexical translation task using a multilayer neural network with comparable performance than classical methods.

This publication has 22 references indexed in Scilit: