Improving the efficiency of distributed deep neural networks and making them fault-tolerant using cognitive computations

نوع: Type: Thesis

مقطع: Segment: PHD

عنوان: Title: Improving the efficiency of distributed deep neural networks and making them fault-tolerant using cognitive computations

ارائه دهنده: Provider: Omid Jamshidi

اساتید راهنما: Supervisors: Mahdi Abbasi, Abbas Ramazani

اساتید مشاور: Advisory Professors: Amir hossein Taherkordi

اساتید ممتحن یا داور: Examining professors or referees: Reza Mohammadi, Alireza Abdollahpouri, Mahlagha Afrasiabi

زمان و تاریخ ارائه: Time and date of presentation: 2026

مکان ارائه: Place of presentation: اتاق سمینار گروه کامپیوتر

چکیده: Abstract: With the growing adoption of deep neural networks in areas such as image recognition, pattern analysis, computer vision, and natural language processing, the need to execute these models in distributed environments has become increasingly important. In such environments, limitations in computational power, high bandwidth requirements, heterogeneous data distributions, and communication latency introduce major challenges and highlight the necessity for efficient, adaptive, and robust approaches. In response to these challenges, this research proposes a novel federated edge computing-based approach to improve fault tolerance and enhance decision-making in distributed deep neural networks. In the proposed method, six end devices are employed for the distributed processing of the MNIST and CIFAR-10 datasets. Each partition of the input data is initially processed at the end devices and the fog layer, where suitable devices are selected for further analysis based on the scores generated by the fog model. The selected outputs are then transmitted to the cloud layer for final refinement, where aggregation methods such as average pooling and max pooling are integrated with cloud computing to improve classification accuracy. By utilizing a cognitive decision-making mechanism in the fog layer, the framework dynamically adapts to network conditions, prediction confidence levels, and the quality of generated outputs. In this architecture, the fog layer evaluates the reliability of device outputs, applies adaptive weighting, and intelligently selects appropriate devices for data transmission to the cloud, thereby implementing behaviors similar to cognitive processes such as analysis, evaluation, and intelligent decision-making. Furthermore, the adaptive selection of reliable devices reduces communication overhead and network traffic. The proposed architecture, referred to as CDDNN, incorporates residual blocks, convolutional layers, batch normalization, activation functions, and fully connected layers implemented across the end devices, fog, and cloud layers. Experimental results demonstrate that the proposed architecture not only improves classification accuracy, but also reduces processing latency, optimizes resource utilization, minimizes the amount of data transmitted to the cloud, and enhances fault tolerance under end-device failure conditions. In addition, the results indicate that the proposed architecture provides stable and reliable performance in the presence of heterogeneous data and dynamic operating conditions. The findings of this research demonstrate the applicability of the proposed framework to Internet of Things (IoT) applications, edge computing environments, intelligent systems, and distributed computing infrastructures.