Work

Reducing the Cloud Overhead and Latency for Artificial Intelligence Applications Using Hybrid Computing

Year: 2021
Type: dissertation
Abstract: In developing artificial intelligence (AI) applications which require high compute resources for training deep learning and machine learning models, cloud is often adopted. Using the cloud in this way offers some benefits, however some limitations exist in terms of increased latency, privacy, reliability and computation overhead on cloud servers. An alternate approach is to run AI-based applications on local devices, but due to limited computing capacity of some local devices certain tasks like training some deep learning neural network models cannot be efficiently done on the everyday local devices, thus requiring the purchase of costly hardware which is not feasible for every user. Therefore, dependence on the cloud sometimes cannot be avoided. To address the limitations with each individual approach (local execution or cloud execution) for certain implementations, hybrid computing is proposed. To reduce the cloud computation cost and latency, a small portion of code is pushed to the client device where the less computationally intensive tasks like deep learning inference is carried out. This concept we refer to as hybrid computing. In this work, a human emotion classification application is developed in a cloud server based solution and a hybrid computing solution. The hybrid computing solution is shown to reduce the cloud overhead and provide better scalability of the application, enabling a larger number of users to utilise the service.
Source: NORMA
Author: Iriogbe, Eromosele Idiagi
Language: English
Open Access status: green
OpenAlex ID: W7052577509
Cites: 19
Cited by:
Related to: 10