Fractal Dimension in Deep Learning

Abstract

In the rapidly evolving field of deep learning, architectural models have grown increasingly complex, delivering impressive performance. However, the more complex models require more processing resources. Furthermore, it requires huge amounts of data to provide high-quality performance results. In this study, we have examined the strengths of the fractal dimension which is a powerful tool for describing self-similarity and complexity of data and for effectively reducing data dimension. Our investigation explores the methods for integrating fractal dimensions into the training of convolutional neural networks (CNNs). We assess this investigation from three key perspectives: performance, training time, and computational resource utilization.