task offloading. The statistical delay constraint is then converted to the number of CPU cycles and constraints on statistical delay using the probability and queue theory. A holistic task offloading algorithm is proposed. The first resource allocation problem is solved to achieve offloading decisions. Efficiency is met by convergence and scalability. An algorithm with these conditions into consideration will provide a solution with much accuracy and efficiency [32].
2.5 Deep Learning Mathematical Models for Edge Computing
Deep Learning is a Machine Learning function of Artificial Intelligence that has been applied in many implementations. Deep learning finds its application in fields that require big data, natural language processing, object recognition and detection and computer vision [33]. Instead of considering explicit data to perform a task, DL uses data representations. The data is arranged in a hierarchy with abstract representations enabling learning of good features [34].
Deep Learning uses cloud computing for performing computational tasks and storage. Latency, Scalability and Privacy were the main challenging concerns of cloud computing, which forced us to choose edge computing over the cloud [33].
Edge computing has solutions for the above challenges of latency, scalability and privacy [33]. Edge computing provides the resource computational tasks at the edge of the devices. The proximity of edge sources to edge end devices is small, which further reduces the edge’s latency. Edge computing works with a hierarchical plan for end devices, edge computes nodes, cloud data centers by providing computing resources at the edge and are scalable to the users. Due to this property, scalability is never an issue. To eliminate any attacks while transferring data, the edge operates very near to the source (trusted edge server), which refrains the data privacy and security attacks [33].
2.5.1 Applications of Deep Learning at the Edge
By providing many solutions, DL finds its vast applications in changing the world. This section will discuss the applications of Deep Learning at the edge [33].
1 i) Computer vision - In computer vision, DL helps in image classification and object detection. These are computer vision tasks required in many fields e.g., video surveillance, object counting, and vehicle detection. Amazon uses DL in Edge for image detection in DeepLens. To reduce latency, image detection is performed locally. Important images of interest are uploaded to the cloud, which further saves bandwidth [33].
2 ii) Natural Language Processing - Speech synthesis, Named entity recognition, Machine translation are a few natural language processing fields where DL utilizes Edge. Alexa from Amazon and Siri from Apple are famous examples of voice assistants [33].
3 iii) Network Functions - Wireless scheduling and Intrusion detection, Network caching are common fields of an edge in network functions [33].
4 iv) Internet of Things - IoT finds its applications in many areas. In every field, analysis is required for communication between IoT devices, the cloud and the user and vice versa. Edge computing is the latest solution for implementing IoT and DL. From much research, DL algorithms are proven to be successful. Examples of IoT using edge include Human activity recognition, Health care monitoring, and Vehicle system [33].
5 v) AR and VR - Augmented Reality and Virtual Reality are the two models where edge provides applications with low latency and bandwidth. DL is considered to be the only pipeline of AR/VR. Object detection is an application of AR/VR [33].
2.5.2 Resource Allocation Using Deep Learning
Allocation of resources optimally to different edge system defines resource allocation. Deep learning uses various learning methods to allocate resources. In this section, deep learning, i.e., the Deep Reinforcement Learning (DRL) allocation method, is discussed. An edge network of green mechanism resource allocation is proposed to satisfy mobile users of their requirements. The “green mechanism” implies increasing the energy efficiency in a system [33].
Table 2.1 illustrates how the methods for allocating resources efficiently with challenges. A DRL method is applied to the edge to overcome this challenge by taking user and base station as requirements. The DRL helps reduce power and bandwidth from the base station to the user, thus making the system energy efficient [33].
The main aim of the DRL method is to provide energy efficiency and a better user experience. Another advantage of DRL is that it has the capability not to exceed the space of the base station. Convex optimization method is first derived to obtain minimum transmission energy and iterate with DQN. It also reduces the space state of the network. On the basis of convex optimization results, optimal connection and optimal power distribution are found. Agent and external environment are the two states of DRL. By taking different actions, the external environment state is achieved. The external environment receives a reward. The main purpose remains as to maximize the value of the reward. In the experimental analysis, several users with three base stations are considered. The number of users for each convergence step is considered. It is seen that as the number of users increased, DRL required more steps for convergence; thus, convergence speed tends to slow down and the efficiency has also increased [33].
Table 2.1 Existing studies using deep learning in edge.
| S. no. | Existing methods | Inference |
|---|---|---|
| 1. | Joint task allocation and Resource allocation with multi-user Wi-Fi. | To minimize the energy consumption at the mobile terminal, a Q-learning algorithm is proposed. In this method, energy efficiency is not considered, which leads to additional costs for the system. |
| 2. | Joint task allocation-Decoupling bandwidth configuration and content source selection. | An algorithm was proposed for avoiding frequent information exchange, which was proven to be less versatile and hence cannot be used in large applications. |
| 3. | Fog computing method for mobile traffic growth and better user experience. | As users are located in different geographical places, implementing fog becomes challenging and requires high maintenance and increased costs. |
| 4. | Deterministic mission arrival scenario | After successfully completing the present mission, each mission is completed, which cannot work as the data source generates tasks continuously, which cannot be handled by the deterministic method. |
| 5. | Random task arrival model | This method works on task arrived and not on the queue tasks, which fails the system to work efficiently. |
2.5.3 Computation Offloading Using Deep Learning
Computation offloading is a great mechanism to offload extensive tasks at the nearby server and communicate cloud with important/filtered data. With edge, computation offloading has excellent applications for mobile devices by enhancing efficiency.
In a study, a dynamic computing offloading mechanism is performed. The objective of the study was to reduce the cost of computational resources. Mobile edge computing is considered (MEC). A Deep Learning method, i.e., Deep Supervised Learning (DSL) is considered. A network of a mobile-based computer system is considered. A pre-calculated offloading
