Dynamic split point computing in multi-task learning implementation with collaborative intelligence

Deep Neural Networks (DNNs) face several challenges in deployment within Internet of Things (IoT) environments, particularly for multi-task robotic and swarm systems. These challenges include hardware limitations, bandwidth constraints, transmission delays, and image packet loss. This thesis proposes a DSPCI-MTL framework under the Collaborative Intelligence (CI) paradigm, which dynamically distributes Multi-Task Learning (MTL)-based computational tasks between edge devices and the cloud. The proposed framework optimizes partitioning points for DNN layers based on real-time bandwidth and data volume while integrating an Autoencoder (AE) architecture to reconstruct lost image packets through feature map similarity for segmentation, classification, and depth estimation tasks. Experimental results demonstrate a 38% reduction in processing time and a 61% improvement in dynamic partition point selection compared to traditional cloud-based methods. The AE-based reconstruction method significantly enhances data integrity in transmission scenarios involving complex and long-range images, providing improved system performance for resource-constrained IoT applications. Additionally, deploying DNNs on resource-limited IoT devices requires balancing computational load and bandwidth efficiency. To achieve Dynamic Partition Point Determination with Metaheuristic Algorithms, the DSPCI-MH approach is introduced. This method adaptively determines optimal DNN partition points between the edge and the cloud by leveraging Grey Wolf Optimization (GWO), Particle Swarm Optimization (PSO), Artificial Bee Colony (ABC), and Artificial Rabbit Optimization (ARO). Unlike static approaches, DSPCI-MH dynamically adapts to network conditions and computational demands, achieving 99.86% faster inference, 99.85% lower energy consumption, and 99.98% improved memory utilization compared to conventional methods. The real-time adaptability of this framework provides a scalable solution for distributed DNN inference, addressing critical challenges in AI-driven IoT systems. The results validate metaheuristic optimization as a transformative strategy for enhancing edge-cloud collaboration in dynamic environments.