Cameras in modern devices such as smartphones, satellites and medical equipment are capable of capturing very high resolution images and videos. Such high-resolution data often need to be processed by deep learning models for cancer detection, automated road navigation, weather prediction, surveillance, optimizing agricultural processes and many other applications. Using high-resolution images and videos as direct inputs for deep learning models creates many challenges due to their high number of parameters, computation cost, inference latency and GPU memory consumption. Simple approaches such as resizing the images to a lower resolution are common in the literature, however, they typically significantly decrease accuracy. Several works in the literature propose better alternatives in order to deal with the challenges of high-resolution data and improve accuracy and speed while complying with hardware limitations and time restrictions. This survey describes such efficient high-resolution deep learning methods, summarizes real-world applications of high-resolution deep learning, and provides comprehensive information about available high-resolution datasets.

Augmented reality (AR) is one of the emerging use cases relying on ultra-reliable and low latency communications (uRLLC). AR service is composed of multiple dependent computational-intensive components. Due to the limited capability of user equipment (UE), it is difficult to meet the stringent latency and reliability requirements of AR service merely by local processing. To solve the problem, it is viable to offload parts of the AR task to the network edge, i.e. mobile edge computing (MEC), which is expected to extend the computing capability of the UE. However, MEC also incurs extra communication latency and errors on wireless channel, therefore it is challenging to make an optimum offload decision. So far, little of state-of-the-art work has considered both latency and reliability of the MEC-enabled AR service. In this paper, we study the scenario multiple edge nodes (ENs) cooperate to complete the AR task. The dependency of task components is modeled by a directed acyclic graph through code partitioning. We aim to minimize the service failure probability (SFP) of the MEC-enabled AR service considering reliability and latency.We design an integer particle swarm optimization (IPSO) based algorithm. Although the solution of IPSO-based algorithm approaches the optimum of the problem, it is infeasible to use IPSO for real-time AR services in practice due to the relatively high computational complexity. Hence, we propose a heuristic algorithm, which achieves performance close to that of IPSO-based algorithm with much lower complexity. Compared with state-of-the-art work, heuristic algorithm can significantly improve the probability to fulfill the targeted SFP in various network conditions. Due to the generic characteristics, the proposed heuristic algorithm is applicable for AR services, as well as for many other use cases in uRLLC.

With the emergence of delay-sensitive task completion, computational offloading becomes increasingly desirable due to the end-user’s limitations in performing computation-intense applications. Interestingly, fog computing enables computational offloading for the end-users towards delay-sensitive task provisioning. In this paper, we study the computational offloading for the multiple tasks with various delay requirements for the end-users, initiated one task at a time in end-user side. In our scenario, the end-user offloads the task data to its primary fog node. However, due to the limited computing resources in fog nodes compared to the remote cloud server, it becomes a challenging issue to entirely process the task data at the primary fog node within the delay deadline imposed by the applications initialized by the end-users. In fact, the primary fog node is mainly responsible for deciding the amount of task data to be offloaded to the secondary fog node and/or remote cloud. Moreover, the computational resource allocation in term of CPU cycles to process each bit of the task data at fog node and transmission resource allocation between a fog node to the remote cloud are also important factors to be considered. We have formulated the above problem as a Quadratically Constraint Quadratic Programming (QCQP) and provided a solution. Our extensive simulation results demonstrate the effectiveness of the proposed offloading scheme under different delay deadlines and traffic intensity levels.

Deploying deep learning models in time-critical applications with limited computational resources, for instance in edge computing systems and IoT networks, is a challenging task that often relies on dynamic inference methods such as early exiting. In this paper, we introduce a novel architecture for early exiting based on the vision transformer architecture, as well as a fine-tuning strategy that significantly increase the accuracy of early exit branches compared to conventional approaches while introducing less overhead. Through extensive experiments on image and audio classification as well as audiovisual crowd counting, we show that our method works for both classification and regression problems, and in both single- and multi-modal settings. Additionally, we introduce a novel method for integrating audio and visual modalities within early exits in audiovisual data analysis, that can lead to a more fine-grained dynamic inference.