Critique a CNN-Based Bridge Safety Proposal
Company: Gsa
Role: Data Scientist
Category: Machine Learning
Difficulty: medium
Interview Round: Technical Screen
A proposed project collects bridge drawings from Wikipedia, trains a CNN to predict whether a bridge will collapse within the first `X` years after construction, and then starts from random image noise and uses gradient ascent on the model output to create a drawing assigned a 99.9% probability of not collapsing.
Critique both the predictive model and the gradient-based drawing procedure. Explain why the optimized image can remain meaningless noise and why the model might systematically predict that old bridges are safe.
### Clarifying Questions to Ask
- How are collapsed bridges represented in the source, and what determines inclusion on Wikipedia?
- What exactly is the label horizon, and are censored bridges handled?
- Which non-drawing variables are available at prediction time?
### Part 1 — Predictive Validity
Identify missing causal variables, label and sampling problems, leakage risks, and an evaluation plan.
#### What This Part Should Cover
- Environment, load, materials, construction quality, maintenance, age, and observation horizon.
- Survivorship and recency or reporting bias in a Wikipedia-derived sample.
- Time-aware splits and uncertainty rather than a single accuracy score.
### Part 2 — Optimizing an Input Drawing
Explain why maximizing a discriminative model's output does not produce a feasible safe design.
#### What This Part Should Cover
- Out-of-distribution optimization, adversarial features, lack of engineering constraints, and model exploitation.
- A constrained validation process if optimization is used at all.
### What a Strong Answer Covers
- A distinction between prediction, causal safety assessment, and design optimization.
- Specific bias mechanisms tied to the proposed data source.
- Safer alternatives and decision limits for the model.
### Follow-up Questions
- How would you represent bridges that have not yet reached the full label horizon?
- What evidence would be needed before using the model in a real review process?
- Could a generative prior solve the random-noise problem by itself?
Quick Answer: A proposed project collects bridge drawings from Wikipedia, trains a CNN to predict whether a bridge will collapse within the first `X` years after construction, and then starts from random image noise and uses gradient ascent on the model output to create a drawing assigned a 99.9% probability of not collapsing. Cover data and labels, leakage-safe features, baselines and model choice, offline evaluation, deployment constraints, monitoring, and drift.