AML Project Planning & Execution 3 — Questions and Answers
Question 1: Which risk is MOST associated with skipping a formal problem framing document before starting ML development?
- Longer training times due to unoptimized code
- Misalignment between the business objective and the ML task formulation, leading to a technically correct but business-irrelevant model (Correct answer)
- Increased cloud infrastructure costs
- Difficulty hiring data scientists
Correct answer: Misalignment between the business objective and the ML task formulation, leading to a technically correct but business-irrelevant model
Without formal problem framing, teams often optimize a proxy metric that doesn't map to the true business goal, producing a model that performs well on paper but delivers no value.
Question 2: What is the purpose of a 'model card' in ML project execution?
- A credit card used to pay for GPU compute resources
- A structured document describing a model's intended use, performance across subgroups, limitations, and ethical considerations (Correct answer)
- A configuration file specifying model hyperparameters
- A Kanban card tracking the model deployment task in Jira
Correct answer: A structured document describing a model's intended use, performance across subgroups, limitations, and ethical considerations
Model cards are standardized documentation artifacts that communicate a model's capabilities, limitations, and responsible use guidelines to downstream stakeholders.
Question 3: During ML project planning, 'technical debt' most commonly accumulates when:
- The team uses Python instead of C++ for model training
- Shortcuts like hard-coded thresholds, undocumented preprocessing steps, and monolithic notebooks are used to hit short-term deadlines (Correct answer)
- The model is deployed on a managed cloud service
- The team holds too many sprint review meetings
Correct answer: Shortcuts like hard-coded thresholds, undocumented preprocessing steps, and monolithic notebooks are used to hit short-term deadlines
Hard-coded values, undocumented transformations, and notebook-first development create fragile, untestable systems that are costly to maintain and extend.
Question 4: A company wants to deploy an ML model under a strict regulatory compliance requirement. Which planning step is UNIQUELY critical compared to a non-regulated deployment?
- Using a GPU cluster instead of CPU training
- Documenting model lineage, maintaining audit trails, and planning for explainability requirements before development begins (Correct answer)
- Increasing the size of the validation set
- Switching from Python to R for statistical compliance
Correct answer: Documenting model lineage, maintaining audit trails, and planning for explainability requirements before development begins
Regulated environments require audit trails, explainability artifacts, and lineage documentation that must be architected into the system from the start, not added retroactively.
Question 5: When planning compute resources for distributed ML training, what is 'communication overhead' and why does it matter?
- Time engineers spend in meetings, which reduces coding productivity
- The time gradient synchronization across workers takes, which can dominate wall-clock training time when the model-to-data ratio is high (Correct answer)
- The cost of sending emails between team members
- Latency in reading training data from disk on a single node
Correct answer: The time gradient synchronization across workers takes, which can dominate wall-clock training time when the model-to-data ratio is high
In distributed training, gradient synchronization overhead can exceed compute time per step, especially for large models on slow interconnects, making network topology a critical planning factor.
Question 6: A team discovers mid-project that the feature engineering pipeline takes 6 hours to run. What is the BEST project execution response?
- Accept the 6-hour cycle time as a fixed constraint and plan experiments accordingly
- Profile the pipeline, identify bottlenecks, and invest in parallelization or caching to reduce iteration cycle time (Correct answer)
- Switch to a simpler model that requires fewer features
- Hire more data engineers to run the pipeline in shifts
Correct answer: Profile the pipeline, identify bottlenecks, and invest in parallelization or caching to reduce iteration cycle time
Long iteration cycles compound across dozens of experiments; profiling and optimizing the pipeline directly reduces total project duration and increases experimentation throughput.
Question 7: Which stakeholder communication practice BEST reduces scope creep in ML projects?
- Avoiding status updates until the model is ready for deployment
- Establishing a formal change request process with documented impact assessment before accepting new requirements (Correct answer)
- Letting stakeholders add features directly to the backlog without review
- Switching from weekly to monthly stakeholder meetings
Correct answer: Establishing a formal change request process with documented impact assessment before accepting new requirements
A formal change request process ensures that new requirements are evaluated for feasibility, cost, and timeline impact before being accepted, preventing uncontrolled scope expansion.
AML Project Planning & Execution 3
Which risk is MOST associated with skipping a formal problem framing document before starting ML development?