AI – model, Data, to Engineering
I once taught a project-oriented master graduation class on artificial intelligence and machine learning. Students were given the flexibility to choose their own projects, and I guided them through the process. This approach, where students learn through new challenges and practical applications, closely mirrors real-world AI practice.
Initially, I focused on teaching the underlying models and theory. Many students, perhaps channeling an academic mindset, focused heavily on the models. While models are fundamental, it’s essential to understand that they are not the sole focus. Some mistakenly believe that the most advanced and sophisticated model is always necessary, overlooking the risk of overfitting that simple models often mitigate more effectively. One must consider if using a complex solution is truly appropriate for the task at hand.
The next critical phase is data. While seeing a model converge is satisfying, the reality of data problems quickly emerges. Issues can arise from cleanliness, diversity, completeness, and various other factors. As the instructor, I pre-screened all project proposals and, in some cases, advised against projects if I anticipated severe data scarcity or other insurmountable data challenges. Struggling with data is a vital part of learning and brings students closer to practical application.
Finally, though less frequently encountered by students in the timeframe of the course, are engineering challenges. Almost all AI/ML products eventually reach this phase. When pushing for higher performance—be it speed, accuracy, or scale—the focus shifts to engineering. This involves dealing with the full technology stack and massive-scale engineering problems, often requiring expertise that goes beyond textbook knowledge. It is gained only through experience and meticulous attention to implementation details. While data remains important, performance optimization requires adjusting numerous other engineering parameters.
