Katkat
Back

Critique a CNN-Based Bridge Safety Proposal

GsaData Scientist · Technical Screen
MediumData ScientistTechnical ScreenMachine Learning

Last updated

This question tests the ability to identify flaws in a computer vision project, focusing on data validity, label sparsity, adversarial robustness, and the practical limitations of predictive modeling in structural engineering.

The Scenario

You are a Data Scientist at a government agency. A research group proposes a bridge safety monitoring system: they plan to scrape bridge architectural drawings from Wikipedia, train a Convolutional Neural Network (CNN) to predict if a bridge will collapse within 20 years, and use gradient ascent on the model input to generate "optimized" bridge designs that the model classifies as highly durable.

Your Task

Critique this proposal. You should address the following areas:

  1. Data & Labeling: Analyze the validity of using public domain images for structural analysis and the feasibility of defining a 'collapse' target variable.
  2. Model Architecture & Baselines: Discuss why a vanilla CNN might be inappropriate for structural engineering and what baseline approaches (physical simulations or heuristic-based models) should be prioritized instead.
  3. Adversarial Risks: Explain why using gradient ascent to generate "optimized" designs is dangerous and susceptible to model exploitation (e.g., adversarial artifacts).
  4. Lifecycle & Deployment: Discuss the risks of model drift, the lack of causal grounding in purely visual patterns, and why this system might fail to generalize to modern construction materials.

Hints

Hint 1

Consider whether an architectural drawing contains the physical metadata (material grade, load tolerances, environmental context) necessary to calculate structural integrity.

Hint 2

Think about the difference between a model's prediction based on correlations in visual data and the actual physical forces involved in a structural failure.

Hint 3

Reflect on how a model optimized via gradient ascent can learn to 'game' the classifier by introducing subtle, non-functional pixel patterns that the model interprets as 'safety'.

Follow-up questions

  • If we were to use visual data for bridge safety, what alternative ML architectures would better respect the hierarchical and geometric constraints of engineering diagrams?
  • How would you measure the uncertainty of your model's predictions, and why is that critical in a high-stakes field like civil infrastructure?
  • If this were a supervised learning problem, how would you handle the extreme class imbalance given that bridge collapses are rare, high-consequence events?