DSPCI-MTL: Dynamic Split Point Computing in Multi-task Learning Implementation with Collaborative Intelligence

Abstract

Deep Neural Networks (DNNs) have become a crucial technology in image processing, renowned for their ability to generate effective feature maps. The integration of DNNs within Internet of Things (IoT) environments, particularly in multi-task robots and swarm systems, has positioned them as vital components in various applications. However, their deployment in IoT devices frequently encounters challenges such as limited hardware capabilities, constrained bandwidth, prolonged data transmission times, and image packet loss due to transmission losses. To address these issues, this paper introduces the Multi-Task Learning (MTL) method of Collaborative Intelligence (CI) strategy by dynamically distributing computational tasks among edge devices and cloud. This method addresses the potential performance degradation caused by suboptimal computational splitting points of DNN for multiple tasks (segmentation, classification, depth estimation) and compensates for losses under varying network conditions and data sizes. A key innovation of our methodology is the introduction of a dynamic method to determine split points by computing DNN layers based on real-time bandwidth and data volume. In addition, an Auto Encoder (AE) architecture is implemented in the cloud to reconstruct image data packets lost during transmission based on feature map similarity measurements. Experimental results show that processing all transactions in the cloud with specific operational parameters reduces processing time by 38 % compared to traditional methods, while dynamically selecting the split point results in gains of up to 61 %. Furthermore, the proposed method achieves efficiency by reducing energy consumption by up to 50 % compared to cloud-only processing. It demonstrates robustness under varying network delays and reduces inference time by up to 47.5 % under low-latency conditions. In this regard, the innovative use of an AE for data loss reconstruction also shows significant potential in complex and long-distance images compared to traditional methods and gives promising results in improving data integrity and system performance. The results confirm the efficacy of the proposed architecture in real-time distributed processing and IoT-based smart systems.