HKU Professor Ping Luo Introduces Innovative DNN Technology for Overcoming Bottleneck in AIoT Devices

Prof. Ping Luo, Assistant Professor, Computer Science, The University of Hong Kong (HKU) visited UMEC today and gave a sharing on the latest deep neural networks (DNN) technology that helps overcome the existing technology bottleneck in AIoT devices.

Prof. Luo was named as one of the 20 honourees of the regional MIT Technology Review Innovators Under 35 Competition in 2020, among a pool of 200 exceptional candidates. He has been working as Research Director in SenseTime Research before joining the Department of Computer Science at HKU. His research focuses on developing Computer Vision and AI technologies to understand human behaviors such as faces, emotions, actions and social relationships, in order to advance human/AI paired systems.

As a key technology bottleneck, the tiny size of AIoT devices has constrained the resources such as memory and computation available for running different AI applications. Accuracy is often scarified to deal with the hardware constraints. Prof. Luo’s dynamic DNN technology aims at maintaining high accuracy under limited memory and computation, by optimising various hyper-parameters, including channels, batch size, bitwidth, branches in a layer automatically. A reparameterised and structural matrix, named Kronecker Reparameterisation is used for dynamic convolution, quantisation and normalisation. It supports complexity-constrained optimisation and improves the accuracy of the DNN.

With the dynamic DNN technology, the performance of AIoT devices can be improved drastically, empowering more edge AI applications. UMEC is interested in combining the technology with its AI chips, systems and applications, in order to deliver a more comprehensive and competent solution to support the intelligent transformation of different industries.

Prof. Ping Luo introduced his research on dynamic, efficient and compact deep neural network

UMEC is interested in applying the dynamic DNN technology to its edge AI chips, systems and applications