Driving AI and Big Data Performance with High-Performance Computing

Dr. Amelie Zhou

Data scientists worldwide have looked for new ways to accommodate the ever-increasing demand for big data and artificial intelligence (AI) in our information-intensive world. Dr. Amelie Zhou, Assistant Professor at the Department of Computer Science, is among the first scholars to explore the use of high-performance computing (HPC) to facilitate deep learning by means of distributed computing. While her technique is already in use in digital entertainment production in other industries, it is also poised to help tackle current social issues.

Humbling Beginning to HPC Research

Dr. Zhou found her research focus in 2011 as she began studying for her PhD at the Nanyang Technological University in Singapore. During this time, she discovered her passion for a particular research focus which was relatively unexplored.

One of the first projects during her PhD work was a scientific workflow study, namely an automatic execution and analysis platform for scientific applications such as earthquake simulation. Dr. Zhou explains: “Researchers have developed accurate models to predict the effects of an earthquake at a particular location, based on factors like how the shockwaves travel, and the quake’s effect on surrounding areas. Our platform can manage the model-based simulation automatically and efficiently. The potential in the application of high-performance computing is endless. This is why I am keen to continue working on it.”

After obtaining her PhD in 2016, Dr. Zhou served as a post-doctoral researcher at the Inria Research Centre in Rennes, France, before taking on an academic role at Shenzhen University in 2017. She subsequently relocated to the HKBU Department of Computer Science in August 2023.

Enhancing Hardware Framework and Optimising Input-output Patterns

According to Dr. Zhou, there are two main pillars in the prevailing studies on HPC technologies. The first encompasses traditional applications in parallel and distributed computing. The second involves modern applications like AI-related model training.

The key emphasis of the first part is the hardware framework, which provides the bedrock for speeding up the processing. Typically, information in an HPC system is handled by computer nodes and is stored in storage nodes. When multiple applications are running concurrently, data transfer between these layers can cause latency. “By adding additional layers in between to enable faster retrieval and saving, we reduce the needed capacity for applications to stop and save data, in turn helping applications to run more smoothly,” says Dr. Zhou.

The next level in HPC involves interference-aware input-output (I/O) scheduling for AI model training. The pattern of data I/O is different from conventional methods as the AI training process commences with a massive amount of input data. While the conventional model handles data in a lower layer for storage, the new process is capable of compartmentalisation and leveraging the network to optimise the I/O execution.

“If we input large volumes of data for the AI training, such as when 100 users are conducting the same function, it is possible to separate them into 10 groups of 10 users, then allocate 10% of resources to each group to complete their tasks. This helps to reduce interference and speed up the process. We are already working on various I/O scheduling scenarios to showcase the power of high-performance computing,” Dr. Zhou explains.

HPC Empowers Digital and Social Entertainment

Over the years, the power of HPC has been witnessed in a variety of industries. The computer-animated film producer Dreamworks has been able to shorten the production time of its popular movie “Frozen” to approximately two years. This is achieved by leveraging the power of HPC. The technique simulated the behaviour of snow and helped with the graphic rendering of snowflakes, rather than using the time-consuming process of rendering each snowflake on a frame-by-frame basis.

Meanwhile, Dr. Zhou reveals that she is collaborating with Meta (formerly Facebook) to optimise their sophisticated deep learning model, which is the backbone of the social media friends’ suggestions and targeted advertising functions. “Meta is capable of recording your user data, analysing your interest in different topics, while serving relevant information to you in real-time. This is a memory-intensive process. We have been working on innovating new hardware other than the central processing unit (CPU) or graphic processing unit (GPU) to alleviate memory constraints and optimise HPC within Meta's model and workload. By delivering improved prediction capabilities and performing pre-processing of user data, the memory requirements for real-time processing can be reduced, leading to enhanced HPC performance,” she elaborates.

Potential Applications for Tackling Social Issues

Further to digital entertainment, the HPC model is applicable to social issues such as the detection of money-laundering and privacy. With AI use in transaction networks in banks and financial institutions, it is possible to detect suspicious activities by tracking a particular user’s unusual transactions either as a sender or recipient.

“Through distributed graph computing techniques, we are capable of tracking transaction connections, rating the level of risk and alerting relevant authorities accordingly. As customer privacy is involved, banks may not be willing to share transaction information, especially the interbank ones. To address this issue, a privacy preserving graph computing is developed to facilitate the collaboration among different banks and to find out the suspicious laundering cases without sacrificing user privacy,” she says.

Community Service Feeds Self-confidence

On top of her busy schedule of teaching, and research as well as being a mother of a young child, Dr. Zhou is actively involved in editing high-impact journals and organising international conferences. These community services help her not only to stay abreast of upcoming trends, but also to seize a stronger understanding of what is being expected in articles and presentations within her field.

Her expertise has been recognised, including with the IEEE Computer Society Technical Community on High-Performance Computing (TCHPC) Early Career Award and the Association of Computing Machinery (ACM) Special Interest Group on High-Performance Computing (SIGHPC) China Rising Star Award in 2021. “I had never applied for awards before 2021. After these wins, I feel so honoured and more confident as a scholar,” says a humble Dr. Zhou.

Reviewing her research career, Dr. Zhou shares her experience: “I would suggest to emerging scholars, particularly female ones, that you should feel confident about yourself and be persistent. By staying focused on the progress of your research, you will succeed in delivering quality research results.”