Edge Computing in Intelligent IoT

Edge Intelligence

Even though the term edge AI or EI is relatively research, new, and methodology in this area have existed for some time. As indicated before, Microsoft developed an edge-based prototype to support mobile voice command recognition, an AI application, in 2009 to show the advantages of edge computing. Despite the early stages of research, EI still lacks a defined definition. EI is currently referred to as the concept of demonstrating AI algorithms on an device with sensor data / signals) which are formed on the device by the majority of businesses and presses. It is important to highlight that, while this description reflects the current most popular tactic to EI in the real world Ike usage of high end AI chips), it significantly reduces the range of possible solutions. (O.salman et al.,2015) Running computation intensive methods locally, as demonstrated by DNN models, is particularly resource-intensive and necessitates the inclusion of high end processors into the device. Such a mandatory requirement in addition to increase the price of EI but is also unsuited and unfavourable to legacy devices that currently exist and have constrained processing power.

This work states that EI is not limited to only using edge servers or devices to execute AI models. In fact, running DNN models with edge cloud collaboration can decrease end-to-end energy usage and latency when compared to the remote execution method, as shown by a numerous recent research. Because of these useful benefits, it is concluded that such a collaborative hierarchy ought to be incorporated into the creation of effective EI solutions. Additionally, current ideas about EI mostly concentrate on the inference phase (i.e., execution the AI model), assuming that the training of the AI model is carried out in the power cloud data centres because the training phase consumes substantially more number of resources than the inference phase. The vast volume of training data must, however, be transported either from end devices or from the cloud, results in excessive communication costs and worries about data protection (O.salman et al.,2015). As a result, it can be seen that EI does not essentially imply that the DNN model is completely trained or inferred at the edge but instead that it can function in a cloud-edge device synchronization manner via data offloading. EI can be categorized into six stages in accordance with the quantity and path length of data offloading, as displayed in Figure. 1. The following list includes a definition for each degree of emotional intelligence.

Figure 1.

Six-level rating for EI

• 1)

  Cloud intelligence: The DNN model is entirely trained and inferred in the cloud.

• 2)

  Level 1—Cloud-Edge Coinference and Cloud Training: Inferring the DNN model using edge-cloud cooperation while training the DNN model in the cloud. In this case, edge-cloud cooperation entails a partial data offload to the cloud.

Level 2—In-Edge Coinference and Cloud Training: This level involves DNN model trainingin the cloud while inferencing it using in-edge technology.The term “in-edge” here refers to the model inference that is being performed within the network edge (using D2D communication).

• 4)

  Level 3—On-Device Inference and Cloud Training: The DNN model is trained in the cloud, but it is inferred entirely locally on the device.

Level 4—Cloud-Edge Cotraining and Inference: The DNN model is inferred and trained through edge-cloud cooperation.

• 6)

  Level 5—All In-Edge: The DNN model is inferred and trained both in-edge.

• 7)

  Level 6—All On-Device: On-device inference and training of the DNN model.