Artificial intelligence (AI) requires large volumes of data to identify patterns and generate predictions rapidly, but questions about how these systems learn, why they fail, and how they can be made more reliable are still not widely understood. Through research at the intersection of machine learning theory and application, Professor Longkai Huang, Assistant Professor at HKBU’s Department of Computer Science, is working to improve AI’s efficiency and accuracy, with a particular focus on the life sciences, where even the smallest detail can make a significant difference.
As an engineering undergraduate at Sun Yat-sen University in Guangzhou, Professor Huang developed a strong interest in automation through his graduation project. “I felt a sense of accomplishment whenever I solved a computational problem. This is why I continued working in the field,” he explains.
After completing his PhD in Computer Science and Engineering at Nanyang Technological University in Singapore, he set his sights on an academic career. However, the disruption caused by the global pandemic meant that few postdoctoral opportunities were available at the time. He therefore moved into industry, joining Tencent AI Lab and Tencent AI for Life Sciences Lab. That experience led to his involvement in AI for the life sciences, including AI-driven research supporting vaccine development for COVID-19 and influenza, as well as work related to rare diseases. It also strengthened his conviction that when a model is used in pharmaceuticals or other safety-critical settings, accuracy and robustness are of utmost importance.
Since joining HKBU in 2025, Professor Huang has found the balance between research and teaching to be a strong fit for him as a young academic. A manageable teaching transition, together with institutional support and the room to build projects steadily, has made it easier for him to focus on his research. Furthermore, HKBU’s collegiate structure means that conversations with colleagues in fields such as physics, chemistry, biology, and Chinese medicine can open up collaborative opportunities across the life sciences.
“In commercial settings, research priorities can shift quickly in response to organisational changes. Academic life offers more room to ask not only whether a method works, but why it works, where it fails, and how it might benefit a wider community,” Professor Huang says.
At the heart of his current research agenda is a fundamental question: how do models actually learn? Modern AI systems contain vast amounts of information, yet the mechanisms by which they organise and use that information remain poorly understood.
For instance, a language model can produce different answers to nearly identical questions depending on how a request is phrased. While this is often addressed through prompt engineering, Professor Huang sees it as a deeper scientific issue. “The inconsistency could be caused by wording alone, limitations in the model’s internal representations, or interactions among multiple, conflicting reasoning pathways. Without a clearer theory, researchers will continue to rely heavily on trial and error. Understanding AI’s capabilities can help us build more efficient, stable, and interpretable models for settings where reliability matters,” he elaborates.
Professor Huang believes that interpretability and efficiency are key to answering these questions. The first concerns how information is encoded and transformed inside machine learning systems; the second asks whether that understanding can reduce the enormous data and computing demands of modern models.
Interpretability matters in particular because models can appear to discover patterns while actually learning from fragile correlations. For example, if a thermometer shows 40 degrees, a model may associate that reading with heat. But if the instrument is faulty or has been tampered with, the conclusion becomes unreliable. “Biomedical data can contain the same kind of misleading signal, which is why reliable models must distinguish meaningful structure from noise and confounding effects,” Professor Huang adds.
This challenge is especially acute in the life sciences, where datasets are limited, and the cost of error is high. In pharmaceuticals, vaccine development, and related biomedical applications, models must generalise well even when training data are scarce or uneven. A better understanding of model behaviour can therefore lead directly to better choices in architecture, data strategy, and benchmarking.
For his work on AI for the life sciences, Professor Huang faces resource constraints, which lead to a slight change in focus. Studying proteins requires costly wet-lab experiments, driving his interest in DNA-level questions. This is facilitated by the emerging area of virtual cell: the use of data-driven models to simulate how cells respond to interventions. From DNA and gene regulation to protein behaviour and cellular response, such systems could help predict drug effects more efficiently and precisely.
Professor Huang sees his work as one part of a fragmented puzzle, much like in other scientific fields, where phenomena appear before the theory that fully explains them.
“AI may now be in a similar phase: researchers can observe behaviour and measure outcomes, but the underlying picture is still incomplete. That makes my work both exciting and humbling, with progress often coming through small experimental clues rather than dramatic breakthroughs,” he shares.
Asked what has remained constant through each stage of this journey, Professor Huang’s answer is immediate: curiosity. That same curiosity began with computing and automation, matured through machine learning, and now drives his efforts to improve AI so that it can optimise resources and contribute to society.
In the long term, Professor Huang believes AI will reach far beyond software efficiency or short-term commercial value. In the life sciences, its most important contributions may lie in helping scientists understand disease mechanisms, identify better therapeutic strategies, and improve the reliability of biomedical decision-making. Whether the goal is safer drug development, more informative cellular models, or future treatments for major diseases, Professor Huang’s underlying aim remains the same: to turn computation into insight, and insight into trustworthy, meaningful human benefit.
