Evaluasi Kelayakan dan Validitas Perbandingan Model Deep Learning Lintas Domain: Studi Kasus YOLO dan RNN

Abstract

The rapid development of deep learning has encouraged the use of various neural network architectures for diverse computational tasks. However, there is a growing tendency to compare the performance of models with different characteristics and objectives without a clear methodological framework, which can lead to scientific misconceptions. This study aims to analyze the validity of a direct comparison between Recurrent Neural Networks (RNN) and You Only Look Once (YOLO). A mixed-method approach was employed, combining a conceptual analysis of fundamental differences including model objectives, data types, output spaces, and evaluation metrics with limited empirical proof within each architecture's respective task domain. The results indicate that RNN and YOLO operate in entirely different representation spaces; RNN is designed to model temporal dependencies in sequential data, whereas YOLO focuses on spatial data processing for object detection. Therefore, it is concluded that a direct comparison between these two architectures is methodologically invalid, as image data lacks meaningful temporal dimensions for RNN processing, and sequential data lacks the spatial annotations required as ground truth for YOLO. Deep learning model evaluation must always be aligned with its original task domain to avoid biased and misleading conclusions.