Gökgöz, Baki; Aydin, Tolga; Gül, Fatih

The traditional Von Neumann computing architecture, which necessitates data transfer between external memory and the processor, incurs significant energy and time costs when running deep learning (DL) and machine learning (ML) architectures. The primary issue with the energy and time efficiency of this architecture stems from the frequent and intensive data transfers between memory and the processor. Therefore, memristive synaptic devices are utilized to overcome this energy and time inefficiency while performing cognitive tasks. The fundamental working principle of memristive devices is to reduce the need for data transfer by combining memory and processing in the same location, thereby significantly decreasing both energy consumption and the time required for operations. However, to achieve the desired level of efficiency in terms of energy and time consumption from neuromorphic systems, the performance of these systems needs to be further improved with respect to accuracy and test error rates for classification applications. Achieving high accuracy performance in such deep learning or machine learning models necessitates optimization processes not only at the hardware level but also at the algorithmic level. In this context, this paper presents a comprehensive examination and comparison of the frequently used SGD and its momentum variants for deep learning and machine learning applications in memristor-based neuromorphic computing systems. The study thoroughly investigates the performance of critical metrics such as the learning properties, energy efficiency, and accuracy rates of the nano-scale titanium dioxide (TiO2) based synaptic device. The experimental results for the MNIST dataset showed AdaDelta 89.48%, AdaGrad 79.00%, Adam 79.13%, AdaMax 79.68%, Momentum 88.55%, Nadam 81.20%, RMSprop 84.91% and SGD 89.47% accuracy. The experimental results for the CIFAR dataset showed AdaDelta 90.51%, AdaGrad 82.08%, Adam 83.10%, AdaMax 81.76%, Momentum 91.25%, Nadam 82.45%, RMSprop 88.11% and SGD 90.21% accuracy. 2024 The Authors. This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License.

Turk, Salih; Bingol, Ozkan; Coskuncay, Ahmet; Aydin, Tolga

X-ray imaging is widely utilized for the detection of bone fractures due to its affordability, rapid processing capabilities, broad accessibility, and ease of use. Despite these advantages, the intricate analysis of X-ray images necessitates advanced computational techniques to fully exploit their rich informational content. Notably, accurate segmentation of these images plays a critical role in aiding medical professionals with precise diagnoses and effective treatment planning. This study examines the impact of integrating different backbone architectures for the task of fracture segmentation in X-ray images. Specifically, the research focuses on enhancing the widely-used DeepLabV3 model by incorporating pre-trained networks such as ResNet50, ResNet101, and MobileNetV3 into the encoder component to improve feature extraction and segmentation accuracy. The FracAtlas dataset, which presents unique challenges due to its small size and the diversity of fractures from various anatomical regions, was employed for model evaluation. Data augmentation techniques were implemented to expand the dataset, and an additional subset focusing on cropped images of fracture areas was developed. The models were trained over 50 epochs, and their performance was assessed using metrics such as Intersection over Union (IoU), loss values, and Dice scores. The results indicate that the DeepLabV3 model with ResNet-based backbones achieved IoU values exceeding 0.93 on the original dataset and demonstrated outstanding performance on the augmented and cropped datasets, with AUC values reaching up to 0.99. The study also highlights the computational complexity of the models, with ResNet101 exhibiting the highest time complexity, while MobileNetV3 was the most efficient in terms of processing time and memory consumption.