This isn't a "toy problem." This is the core problem for autonomous driving.
Think about the analogy:
"Synthetic clock" = A perfect, forward-facing stop sign, in bright sunlight, as seen in the training data.
"Dali-esque warped clock" = A real-world stop sign. It's tilted, partially occluded by a tree branch, covered in graffiti, and seen at a 30-degree angle during a foggy sunset.
This study shows the AI doesn't just get the "time" wrong; it has a "cascading failure" where it can't even identify the hands properly. For a car, this means it doesn't just "misread" the sign, it might fail to see the sign as a sign at all.
It's a perfect example of the gap between "in-distribution" performance and "out-of-distribution" robustness.
This isn't a "toy problem." This is the core problem for autonomous driving.
Think about the analogy:
"Synthetic clock" = A perfect, forward-facing stop sign, in bright sunlight, as seen in the training data.
"Dali-esque warped clock" = A real-world stop sign. It's tilted, partially occluded by a tree branch, covered in graffiti, and seen at a 30-degree angle during a foggy sunset.
This study shows the AI doesn't just get the "time" wrong; it has a "cascading failure" where it can't even identify the hands properly. For a car, this means it doesn't just "misread" the sign, it might fail to see the sign as a sign at all.
It's a perfect example of the gap between "in-distribution" performance and "out-of-distribution" robustness.
[dead]