Computers have long been able to ‘see’ through cameras, but they’re increasingly being taught to understand what they’re seeing in the same way we do. Advances in artificial intelligence (AI) are boosting today’s computer vision technologies, allowing them to lend us extra eyes in cancer cell detection, crop health monitoring and other critical tasks.
However, AI can be limited by its training data. Computer vision models generally learn from the images they’re fed with and the labels attached to those images, but have trouble applying their knowledge to slightly different situations. For example, a model trained to identify pedestrians in photos taken in bright daylight might be confused by photos of rainy streets. This is where the human eye has an advantage—the ability to seamlessly adapt to different environmental conditions. In computer models, this ability to adjust to new contexts is known as domain adaptation.
Researchers like A*STAR Graduate Scholar Mohamed Mikhail Kennerley are working on domain adaptation in computer vision models, hoping to boost their reliability in the variable conditions of real-world settings. Drawn to computer vision by its diverse applications and real-world impact, Kennerley is pursuing his PhD degree at the National University of Singapore (NUS) and the A*STAR Institute for Infocomm Research (A*STAR I2R).
In this interview with A*STAR Research, Kennerley shares his journey in programming, his perspectives on computer vision research, and how he hopes to inspire a new generation of scientists.
1. What drew you to science and computing?
I’ve always had a strong curiosity about things I could tinker with, and an urge to do so—whether to solve problems or even create new ones, just to figure out how to fix them. My first exposure to programming itself was in Temasek Polytechnic, where I took a module on C++. However, I only started to enjoy programming during my undergraduate studies at the Singapore Institute of Technology, where we used MATLAB to solve optimisation and pathfinding problems in robotics.
My first job was as an engineer, but rather than remain in that field, I wanted to deepen my knowledge of computing; so I picked up Python through online courses while working. This led to my pursuing a part-time master’s degree in intelligent systems at the NUS Institute of Systems Science (NUS-ISS), which had a good balance of research and applied problem-solving.
My master’s project revolved around vision-based problems, which showed me how vast and impactful computer vision research can be. That experience, as well as a drive to have more ownership of my work, led me to apply to the A*STAR Graduate Scholarship towards a PhD degree in that field.
What excites me most about computer vision is how we can teach computers to perceive and interpret visual data the way humans do by instinct. Human vision is incredibly robust: we can recognise objects in different lighting conditions, from different angles, and even in incomplete or noisy (distorted) images. Replicating that level of adaptability and understanding in machines remains a huge challenge. AI has helped advance the field significantly, but there are still gaps in robustness and generalisation that make it compelling to me.
Beyond the challenges of the research itself, computer vision is also making incredible impacts in everyday life, from autonomous self-driving vehicles to the facial recognition systems you use to unlock your phone. Yet we’re only scratching the surface of what’s possible; we continue to expand the capacity of machines to understand the world, and how to apply that capacity in new ways.
2. Tell us about your current work.
My current research focuses on domain adaptation; specifically on improving how computer vision models transfer knowledge across different environments. A well-known use case for this is in autonomous vehicles, as cars must operate in diverse and unpredictable conditions: if a self-driving car is trained mostly on clear weather data, it might struggle in fog, rain or low light. Vision models that can reliably generalise across such variations have broader applications beyond vehicles, extending to areas like medical imaging, surveillance and automation.
In my current research on night-time object detection, we focus on a semi-supervised machine learning technique called student-teacher training, where a ‘student’ model learns from a ‘teacher’ model. We initially train the student with labelled day-time data, then copy its model weights—the numbers it uses to make predictions—to the teacher, which updates its own weights based on the student’s work. With several training iterations, the teacher resembles the average of many previous students, making it a more robust model.
Once this is set up, we have the teacher generate pseudo-labels for unlabelled night-time images, which the student then learns from. The teacher is updated with the student’s results before repeating the process. The result is a student-teacher model that can perform well in both day and night scenes.
One of the biggest difficulties in domain adaptation is that many existing methodologies rely on similar core ideas, which are not universally effective. Newer approaches may introduce fresh concepts but often fail to perform consistently in practical settings. My research explores striking the right balance between improving generalisation and maintaining usability.
3. What are some other interesting projects you work on?
Beyond research, I am actively involved in Mendaki Club, a community of young Malay/Muslim professionals. Partnering the Mendaki Tuition Scheme, one of the programmes we run introduces students to careers and industries they might not have considered otherwise, broadening their perspectives on future opportunities. I believe that representation and exposure can make a real difference in their aspirations.
Within this programme, we conduct a miniworkshop on computer vision where we introduce students to the basics of AI. They get a guided hands-on experience of building a simple classification model, which emphasises how diverse and high-quality data affect a model’s performance. The goal is to make AI approachable, showing students how it applies to real-world problems.
Through this initiative, I’ve realised the importance of demystifying AI for young students who often see it as abstract or overly complex. Making these concepts more accessible can inspire the next generation to get more involved in research.
4. What excites you about your future in research?
I think it’s the fact that research is dynamic; I’m not confined to the topic I’m currently pursuing but am able to pivot and explore new problems in different areas. While I enjoy working on domain adaptation, I also see opportunities beyond this area that I might explore in the future. There is always something new to learn, and often, the journey is more important than the destination.
Ultimately, I want my work to have an impact on society. Whether that means improving AI’s reliability and accessibility, or applying it to critical areas like healthcare, I want my research to contribute to real-world progress. I believe in research done not just for innovation’s sake, but for the creation of meaningful change that benefits people.
5. What advice do you have for young aspiring researchers?
The common advice is to have a passion for research, but if you’re already considering this path, you likely already have that passion. What’s more important is to be okay with failure. Research is filled with trial and error; many ideas that seem promising will not work the first time or even the tenth time. It’s easy to get frustrated, but learning from failure is what drives progress.
If you decide to pursue research, go in with the mindset that failure is not the end, but rather a stepping stone toward discovery. Stay curious, keep asking questions, and surround yourself with mentors and peers who challenge and inspire you. The journey is tough, but when you finally reach the answers to the questions that drive you, it’s incredibly rewarding.
