DEVELOPMENT OF A SCALABLE AI PLATFORM BASED ON INTEGRATION OF EDGE COMPUTING WITH CLOUD TECHNOLOGIES

Abstract

Relevance of the research topic. Recent trends show that AI is becoming more widespread in all areas of life. One such broad area is edge computing, and how to best leverage their strengths and weaknesses to deliver quality services to the end user. Today's cloud systems use a large number of data centers located around the globe to increase performance and improve latency, but they often do not offer flexibility in terms of user experience for both users and the companies that operate them. 

In order to get an understanding of how much the system can be used as accurately as possible, we introduce several metrics that will help to obtain system parameters by which the proposed system can be compared with alternative solutions.

Formulation of the problem. The deployment of AI software should be accompanied by a thorough analysis of the target system. This article will explore different approaches to deploying AI in edge systems. One will include a smaller model suitable for low-power devices, and the other will be a full-scale model. Several tests will be conducted to measure the timing and accuracy of the proposed system.

Analysis of recent research and publications. Scientific research in areas related to increasing the speed and accuracy of delivery of AI-enabled systems is not widely explored. Edge and Fog computing technologies offer a wide range of performance optimizations, but little research has been published on running AI models on Edge/Fog platforms.

Setting objectives. The goal of the paper is to determine the feasibility of integrating edge computing with cloud computing and see how well it performs with AI tasks.

Presentation of the main material. Modern solutions, as a rule, involve the use of autonomous IoT devices [1] or cloud solutions [2]. Edge computing is a relatively new model of device networking, creating demand for solutions capable of rapid deployment and capacity scaling. Most research efforts are focused on building a more powerful AI model or scaling down the model to try to run it on an IoT device. Small current research includes using the cloud to facilitate the penetration of artificial intelligence and exploring ways to improve the quality of service. The resulting system can be applied in a variety of existing industries, in order to obtain significant economic and social benefits at a relatively low cost of implementation.

Conclusions. This research shows a way to improve the performance of embedded artificial intelligence on IoT devices using an Edge architecture. To illustrate this, a system was developed that demonstrated the ability to rapidly deliver ML results over a network. Using cloud computing resources, we were able to deploy such a system much faster than using a conventional on-premise cluster, but this approach is not limited to the cloud. There is room for improvement. The accuracy of DNNs can be improved by various means, for example, further tuning of the hyperparameters or changes to the architecture can bring benefits. Using RNNs can be useful for time-sensitive data, or other datasets can be used to improve accuracy by augmenting the original dataset. All this can be easily done in the cloud. This research proves that we can extend advanced computing capabilities with cloud services.

Unexplored aspects. The reliability of this solution can be related to its architecture. It is worth investigating the proposed method with different settings, for example: increasing the number of EC2 instances to ensure a larger data throughput, creating a separate service for data management. The regional infrastructure solution presented in this publication can be improved by adjusting the parameters and topological organization of the network. Perhaps more cloud services can be hosted on the client side to reduce costs. Finally, additional studies should be conducted to determine the financial benefits of the system obtained by this method.

Keywords: AI, clouds, scaling.