Derin Öğrenme Modellerinin Doğruluk, Süre ve Boyut Temelli Ödünleşme Değerlendirmesi

Abstract

Machine Learning and especially deep learning models need to be optimized over three main criteria concurrently, to be operationalized in real-time field applications. These criteria are model’s accuracy, training-testing times and file size. Related work only considers two criteria (e.g. accuracy-time) together. However, it is observed that deep neural networks (DNN) designed to improve model accuracy can increase training time and size, while efforts to reduce model size can lead to lower accuracy. A trade-off needs to be made among these three criteria. In this paper, to demonstrate the effects of different optimization techniques on model performance, we tested ResNet50, ResNet101, VGG16, VGG19, EfficientNet pre-trained models with CIFAR10, CIFAR100 image datasets, which are commonly utilized in the DNN research field. Important performance results obtained over Google Colab Pro and TensorFlow system show that weight quantization is the most successful technique so far in multi-dimensional optimization, while weight clustering and transfer learning techniques remain useful in 2-dimensions. In addition, we designed and tested a new DNN operational score and model-to-model layer transfer method for the first time in literature. We hope that our framework will constitute a multi-dimensional evaluation reference for DNN models before they are operationalized.