Project PI : Ngai Man (Man) Cheung
Project Period : 01 December 2020 to 30 November 2022
Deep Neural Networks (DNNs) have been playing a crucial role in the recent tremendous improvements of many artificial intelligence and computer vision tasks, including image classification, object detection, segmentation and image retrieval. However, these improvements come at the cost of requiring an enormous amount of memory and computational resources. These impede the applications of DNNs on resource-constrained platforms such as embedded devices and intelligent surveillance cameras. Therefore, new methods to make DNNs become compact, lightweight and efficient in terms of memory and computation are important to enable DNNs on many resource-constrained platforms and achieve edge intelligence.
The objective of this research is to investigate new learning algorithms and network architecture designs for extremely compact deep neural networks (DNNs): DNNs that have very low power, memory and computation requirements. These compact DNNs can be efficiently implemented on resource-constrained platforms. The systems can be deployed in battlefield for situation analysis and monitoring. Importantly, compact DNNs enable analytics and processing on the edge devices, removing the reliance on stable communication channel between the edge devices and backend servers, which may not be available in adversarial situations.
Potential applications: Our research findings shall enable extremely compact DNNs to be used on power-efficient FPGA or other embedded platforms. This enables resource-efficient embedded analytics systems that can perform various AI tasks at the edge: anomalous event detection, urban surveillance, incident detection, etc. The system can be used in autonomous robot to carry out military tasks in battlefield where connectivity can be poor and data needed be analyzed on the device without delay.
Project team expertise: Image Processing, Computer Vision and AI on resource constrained platforms; AI with limited data