Redes neurais artificiais adaptativas: abordagem evolutiva para melhoria de performance e eficiência

Abstract

In recent years, artificial neural networks have emerged as one of the most powerful tools in the field of artificial intelligence, offering innovative solutions across a variety of applications. However, the development of these networks often faces significant challenges related to computational efficiency and environmental impact. This research aims to investigate these challenges, particularly in the process of creating artificial neural networks, and propose a solution to enhance their efficiency in terms of predictive quality, hardware constraints, and environmental impact reduction. Building on the hypothesis that adaptive activation functions are capable of creating more complex neurons in the architecture, thereby reducing the need for deep learning approaches that result in heavier and more complex models, the primary focus of this study is to propose the development of an evolutionary strategy for creating artificial neural networks using adaptive activation functions. The objective is to find models with higher predictive power and the least possible complexity. This endeavor seeks to contribute to the reduction of carbon dioxide emissions and other gases associated with climate change resulting from the execution of these models. Four sets of numerical experiments were conducted to evaluate the hypothesis and the proposed solution. The first two experiments compared the efficiency of adaptive activation functions against traditional activation functions in classification and regression problems. The third experiment combined evolutionary approaches with adaptive activation functions to find the best model in terms of metrics and complexity in Autoencoder problems. Finally, in the fourth experiment, the application of evolutionary approaches combined with adaptive activation functions was extended, this time focusing on optimizing the model in terms of metrics and complexity in a real context of brain tumor detection. All obtained results demonstrated a gain when using Adaptive Activation Functions in architecture construction, along with an environmental impact comparable to traditional strategies. These findings represent a significant advancement in the pursuit of more effective and environmentally conscious artificial neural networks, aligning Evolutionary Optimization with environmental responsibility, especially in reducing Carbon Footprint.