Most engineers treat the loss function as a purely technical choice but in safety critical systems like ADAS, it’s actually an operational one.
Take decision trees as a example. They split data by reducing impurity, and measuring how mixed outcomes are in each branch to create cleaner, are uniform groups. It's good on paper but mathematically clean splits can still be operationally dangerous which are called overfitting.
If a model treats a pedestrian stepping into the road the same as one on the sidewalk, the numbers might look okay. But in a real vehicle driving scenario, that's exactly the kind of guess you can't take.
The real question is not "Does this split lower the impurity score?"
It is "Does this split actually separate the scenarios the car needs to handle differently?"
In ADAS, a false negative on a pedestrian doesn't carry the same cost as a false negative on a stationary cone. A standard impurity metric doesn't know the actual difference. It just sees numbers.
Your loss function has to reflect the real-world cost structure of your problem, not just not just make the training chart look pretty.
How do you handle cost-sensitive decisions in your models?
Top comments (0)