Electrical and Electronic Engineering

The Artificial Intelligence of Things (AIoT) is the combination of Artificial intelligence (AI) with the Internet of things (IoT) to enable autonomous decision-making, data analytics, and system optimization. AIoT for smart cities allows the collection of enormous sensor data for a better understanding of the environment, human behaviors, and city operations, which leads to more efficient resource management and promotes a sustainable and healthier society. The Smart Water Auditing project aims to use IoT and machine learning to provide insights into how water is being used in the households of Hong Kong to reduce the consumption of water and raise awareness of people’s water consumption habits. Our talk will present our design workflow, IoT infrastructures, machine learning algorithms, and experimentation for water end-use disaggregation in Hong Kong.

Human brain can perform many tasks much better than classical electronic computers, such as face recognition, reasoning based on vague information, and learning from experience, to name a few. Recently, brain-inspired algorithms have promoted in the rapid development of artificial intelligence, however, they cannot work well in classical computers. In this talk, Dr. Can Li will present his recent works on building brain-inspired computers to fit better with brain-inspired algorithms. Those computers are based on an emerging nanoelectronics device – a memristor – which can store information and compute simultaneously, similar to synapses and neurons in our brain. The built hardware can function similar to human brains, for example, it can tolerate hardware defects, make full use of the nonlinearity of devices, learn from rare samples, and so on.

Machine learning and deep neural networks have revolutionized various fields, most obvious examples are computer vision and natural language processing. Apart from the surging sizes of sophisticated models, an emerging trend is to go down the opposite route of deploying lightweight models on the edge (terminal or user end) for relatively simple AI tasks. This is named edge AI which is often constrained to run under restrictive compute and storage resources. In this talk, we will explore the latest theory in neural network modeling that allows the total avoidance of AI training that used to be slow, daunting or even impossible for the edge. Specifically, we will scratch the surface of the neural tangent kernel, and try to establish (well…. qualitatively) the equivalence of data and network, such that once the data are ready, the network is instantly ready, too.