Prof. Zhou has been working on understanding deep neural networks and making Artificial Intelligence (AI) models more trustworthy. He received his master’s degree in information engineering at CUHK in 2012, and later obtained his PhD at Massachusetts Institute of Technology (MIT) and returned to CUHK in 2018. MIT Technology Review recognised him for making AI models more understandable and trustworthy to humans.
Unpack AI “black-box” and make AI models more understandable and transparent
AI models such as deep neural networks are increasingly employed to make highly consequential decisions in daily life, including auto piloting, credit assessment before granting a loan, and facial recognition. However, these complex models, purely trained on massive amounts of data, are often treated as “black-box” due to their opaque and complicated internal processing mechanism, which cannot be understood by ordinary people and even professional engineers. This “black-box” property of AI models raises serious safety and reliability concerns that greatly limit their applicability.
In view of these issues, Prof. Zhou has made significant contributions in interpreting the network’s representation and output prediction, by developing innovative techniques such as Class Activation Mapping and Network Dissection. These techniques help researchers and practitioners better explain the model prediction and diagnose the mistakes made by the AI models, which can be applied to fields such as autonomous driving, medical image diagnosis, and healthcare.
Recently, his research team has developed interpretation methods to discover the knowledge learned by deep generative networks for image synthesis. It can achieve realistic photo editing for faces and scenes, such as changing the age of a person and improving the lighting condition of a selfie photo. In his research collaboration with MIT-IBM Watson AI Lab, he and his team visualize an AI model’s blind spots, a new tool that reveals what AI models leave out in creating a scene. With a focus on enabling machines to sense and reason about the environment with more interpretable representations, he hopes researchers will pay much more attention to characterizing transparency and interpretability of the models ignored in the machine-learning systems.
Prof. Zhou remarked, “I would like to acknowledge the huge amount of support from my former mentors and collaborators at MIT, as well as my current research group at CUHK. I am confident that we will continue leading the frontier for developing interpretable, robust, and safe AI technologies. In particular, this year is the 30th anniversary of our Department of Information Engineering at CUHK, I hope there will be more and more of the younger generation in Hong Kong who will pursue a career in Artificial Intelligence and Information Technology.”
About “Innovators Under 35” Asia Pacific
“Innovators Under 35” Asia Pacific has been a prestigious regional recognition from MIT Technology Review since 1999. It is a list of technologists and scientists, all under the age of 35, whose work is changing the world. It recognises the development of new technology or the creative application of existing technologies to solve global problems in industries such as biomedicine, computing, communications, energy, advance materials, software, transportation, web and internet. The Asia-Pacific list covers countries and regions in South-east Asia, Australia, New Zealand, Hong Kong and Taiwan.
About Information Engineering, CUHK
The Department of Information Engineering was established by Prof. Charles Kao, who was the Vice-Chancellor of CUHK in 1989. It was the first and is still the only Information Engineering Department in Hong Kong. It has gained a world-wide reputation for its leading edge research and top quality programmes. The Department focuses on areas related to generation, communication and networking, storage, signal processing and security of information involved in many real world applications. It contains areas that are traditionally linked to computer science like internet design and computer networks, machine learning and artificial intelligence, big data analytics, cyber security, mobile and cloud computing, as well as areas that are more traditionally linked to communications aspects of electrical engineering such as optical and wireless communications, multimedia processing, communication theory and applications of the internet of things.
Website of MIT Technology Review “Innovators under 35”: https://www.innovatorsunder35.com/the-list/bolei-zhou/