The question I kept asking my AI students
Mentoring work on imagery and radar kept bringing me back to a simple question: have we tested the situations in which the model might be wrong?
There is a delicate moment in a student project when the model finally works. Weeks of effort have turned into a result the student can show. As a mentor, I want to give that achievement its due, then open the next question without taking the excitement out of it: what happens when the next example is different?
That balance was part of my industry mentoring in the SAYZEK Academic Thesis Programme in 2024–2025. The students were working on AI for UAV imagery and on localisation and classification from radar data. The work gave us plenty to discuss beyond a single score: how a measurement was made, what the model had seen before, and what its result would mean somewhere new.
When a test looks a little too familiar
An image can look new to us while remaining very familiar to the model. A different file may come from the same flight, sensor or lighting conditions as the training data. Our conversations stayed with these general research questions; student identities and restricted material are not part of this account.
A training set and a test set can contain different files but very similar circumstances: the same sensor, lighting or flight. A strong score then tells us something useful about those circumstances. It may tell us less about a new place or an unfamiliar sensor.
The question that opened the conversation
“When might this model be wrong, and is that situation in the test?” The answers made the next experiment more concrete. For imagery, it might be low sun, partial occlusion or an unfamiliar viewpoint. Rare classes also deserved a closer look, because a good overall score could hide how poorly they were recognised.
Radar needed its own reasoning. Clutter and multipath are features of how signals meet an environment; treating everything difficult as disposable noise can remove the very conditions the model needs to understand. The useful lesson was to design the test around the setting, rather than borrow it unchanged from another task.
What stayed with me afterwards
The habit travelled into my QANTIS work: say where a result holds, try the edge of that range and keep the failures in the account. It can feel uncomfortable to describe where a method degrades. For another researcher, that description may be the most useful part of the paper.
Between 2024 and 2025 I was an industry mentor in the SAYZEK Academic Thesis Programme, run by Türkiye’s Presidency of Defence Industries and the Council of Higher Education. The programme received 421 applications from 119 universities, accepted 356 and involved 90 mentors from 20 companies. Students submitted three interim reports for review.
My certificate, dated 10 July 2025 and signed by the President of Defence Industries, records the mentoring contribution. I value it for that experience. The most lasting part was learning how much a carefully phrased question can change a research conversation.
- Find the unfamiliar case
- Keep it separate
- Read the errors together