GSA Data Scientist Interview Experience — A CNN Adversarial-Example Question on Bridge Collapse Prediction

We talked about research.

Explain why this approach would fail: collect bridges and their design drawings from Wikipedia. Train a CNN on these drawings to predict whether a bridge will collapse (or whether it will collapse within the first X years after being built). Then, starting from random noise, use gradient descent to "construct" a drawing plan that makes the model predict a 99.9% probability of "not collapsing."

The resulting drawing plan would just be random noise.

Explain why the prediction model would fail:

Missing other key features, like weather/environmental conditions, bridge load, year built, etc.

We also got into earthquakes and horses, not sure why — that guy was kind of out there.

Give a possible reason why the model might keep predicting that old bridges won't collapse:

The dataset has survivorship bias — older bridges still in use are more likely to show up on Wikipedia (collapsed ones don't).

Recency bias (not sure if that's exactly the right term) — recently collapsed newer bridges are also more likely to be included on Wikipedia.