5 Questions to Test Your ML Model Development Knowledge (MLA-C01)
Here are 5 MLA-C01 practice questions on ML Model Development. It makes up about 26% of the exam. Each answer explains the correct choice and what is wrong with the others.
AWS Machine Learning Engineer – Associate (MLA-C01) · ML Model Development · 5 questions with answers and explanations
🧠Reading ML Model Development questions
Model development questions give a modeling goal plus a constraint (no labels, thousands of related series, a training run that stalls, versions that need approval) and ask which SageMaker capability or algorithm fits. Most options are real SageMaker features, so the work is matching each requirement to the feature built for it and ruling out tools that belong to a later stage.
Keep in mind
• Decide whether labeled data exists before looking at algorithms. Finding unusual points without labels calls for an unsupervised method; supervised options assume labeled examples.
• Place each tool in the lifecycle: some inspect a training job while it runs, others watch an endpoint after deployment. A problem seen during training needs a training-time tool.
• For forecasting across many related products, including new ones with little history, think of one model trained across all series that can also give ranges, not one model per item.
📊 Set profile
Exam: MLA-C01 · 65 questions · 130 minutes · passing score 720 · the second largest of 4 domains
This set: 4 advanced, 1 applied · 3 scenario, 2 concept check
Topics: Amazon SageMaker, SageMaker DeepAR, SageMaker JumpStart, Amazon SageMaker Model Registry, Amazon SageMaker AI
Question 1 of 5
During training of a very deep neural network, a data scientist observes that the loss stops decreasing early and suspects that gradients in the earlier layers have become extremely small. Which SageMaker capability can automatically monitor tensors during training and trigger an alert when this specific condition (vanishing gradients) is detected?
Show answer
✅ Answer: A
Difficulty: advanced · Topic: Amazon SageMaker
Vanishing gradients
In a deep network, gradients are propagated backward layer by layer. With some activations and initializations they are multiplied by small numbers again and again, so the earliest layers receive gradients close to zero. Those layers stop learning and the loss plateaus early.
SageMaker Debugger
Debugger saves tensors (weights, gradients, losses) from a training job at intervals and evaluates rules against them in parallel with training.
• Built-in rules include vanishing_gradient, exploding_tensor, overfit, loss_not_decreasing and poor_weight_initialization.
• A triggered rule changes status and can raise an Amazon CloudWatch event, notify the team, or use a built-in action to stop the training job.
Example: the vanishing_gradient rule flags when the mean absolute gradient falls below a threshold such as 1e-7.
Why each option works or fails
• ✓ A. Debugger's vanishing gradient rule (correct): Debugger monitors gradient tensors during training and its built-in rule detects exactly this condition and can trigger an alert.
• ✗ B. Clarify bias detection: Clarify measures data and prediction bias and feature attributions; it does not watch gradients during training.
• ✗ C. Model Monitor data quality: Model Monitor checks data captured from a deployed endpoint against a baseline; it runs after deployment, not inside a training job.
• ✗ D. Ground Truth active learning: this reduces labeling effort by auto-labeling confident examples; it has no view into a model's training tensors.
🎯 Key idea: Problems inside a running training job, such as vanishing gradients, are caught by Debugger rules; Model Monitor watches deployed endpoints.
Question 2 of 5
A retailer must forecast demand for 20,000 related products, including new products with little history, and provide prediction intervals for inventory planning. Which approach fits?
Show answer
✅ Answer: A
Difficulty: advanced · Topic: SageMaker DeepAR
Forecasting many related series
Classical methods fit one model per series. With 20,000 products, many with little history, that works poorly: a new product has too few points to learn from.
DeepAR
SageMaker's built-in DeepAR is a recurrent neural network trained on all series together. It learns patterns shared across products (seasonality, promotions, trends), so a new product can borrow strength from similar items, especially when category features are supplied.
It is probabilistic: instead of a single number it predicts a distribution, returned as quantiles such as p10, p50 and p90.
Example: for next week a product forecast might be p10 = 80, p50 = 120, p90 = 190 units. Planners can stock to the p90 for key items to avoid stockouts, or to p50 where overstock is costly.
Why each option works or fails
• ✓ A. DeepAR with quantile forecasts (correct): one global model learns across 20,000 related products, handles items with short history, and returns prediction intervals for inventory planning.
• ✗ B. K-means on sales: clustering groups similar products but produces no forecast and no uncertainty range.
• ✗ C. XGBoost point estimates: a single predicted value per product gives no prediction interval, which the planners explicitly need.
• ✗ D. Linear regression per product: 20,000 separate models cannot share information, so products with little history get unreliable fits, and it adds a large maintenance burden.
🎯 Key idea: Many related series, cold-start items and a need for intervals point to a global probabilistic forecaster like DeepAR, not per-item or point-estimate models.
Question 3 of 5
A team wants to try several pretrained open models for text classification, fine-tune one on its data, and deploy it with one click, in its own account. Which capability fits?
Show answer
✅ Answer: C
Difficulty: advanced · Topic: SageMaker JumpStart
SageMaker JumpStart
SageMaker JumpStart is a model hub inside SageMaker. It lists hundreds of pre-trained models, both open-weight models (for example from Hugging Face and other providers) and proprietary ones, grouped by task such as text classification, summarization or image classification.
For each model JumpStart offers:
• Deploy: one click (or one SDK call) creates a SageMaker endpoint in your own account.
• Fine-tune: point it at your labeled data in S3 and it launches a training job with sensible default hyperparameters.
• Example notebooks that show evaluation and inference.
So a team can deploy three candidate text classifiers, compare them on a test set, fine-tune the best one on its own tickets and deploy the result, with everything running on its own SageMaker resources.
Why each option works or fails
• ✗ A. Feature Store: Feature Store stores and serves engineered features for training and inference; it holds no pre-trained models.
• ✗ B. Ground Truth: Ground Truth is for labeling data with human workers. It can help create training labels, but it does not provide or deploy models.
• ✓ C. SageMaker JumpStart (correct): it provides pre-trained open models to try, built-in fine-tuning on your data and one-click deployment, all in your own account.
• ✗ D. SageMaker Neo: Neo compiles an already trained model for efficient inference on specific hardware; it does not offer a catalog of models to try or fine-tune.
🎯 Key idea: A catalog of pre-trained models with fine-tuning and one-click deployment is JumpStart; Feature Store, Ground Truth and Neo handle features, labels and compilation.
Question 4 of 5
A team wants to compare several trained model versions and approve the best one for deployment with lineage tracking. Which feature fits?
Show answer
✅ Answer: C
Difficulty: advanced · Topic: Amazon SageMaker Model Registry
Governing model versions
As a team retrains, it accumulates many model versions. To deploy safely it needs to know which version is which, how each was built and who approved it.
SageMaker Model Registry provides this:
• A model package group collects all versions of one model, for example churn-model. Each registered model package is a numbered version.
• Each version stores the container image, model artifact location, evaluation metrics and optional metadata.
• An approval status (PendingManualApproval, Approved, Rejected) records the decision; an approval can trigger a CI/CD deployment, for example through Amazon EventBridge.
• Lineage tracking links a version back to the training job, data and pipeline that produced it.
Why each option works or fails
• ✗ A. Email votes: the decision lives in inboxes with no link to the exact artifact, metrics or training data, so there is no auditable lineage or approval workflow.
• ✗ B. Overwriting one S3 key: each upload replaces the previous model, so older versions are lost unless versioning is enabled, and nothing records metrics, lineage or approval.
• ✓ C. Model Registry with package groups and approval status (correct): it versions every model, keeps metrics side by side for comparison, records lineage, and captures an explicit approval before deployment.
• ✗ D. Spreadsheet of S3 paths: it is manually maintained, can drift from reality, and has no enforced approval state or automatic lineage to training jobs.
🎯 Key idea: Versioning, comparison, lineage and deployment approval belong in SageMaker Model Registry, not in ad hoc lists, emails or overwritten files.
Question 5 of 5
A company wants to detect unusual spikes in server metrics without labeled anomaly data. Which algorithm fits?
Show answer
✅ Answer: B
Difficulty: applied · Topic: Amazon SageMaker AI
Anomaly detection without labels
Labeled anomaly data is rare: spikes are infrequent and nobody has tagged them. An unsupervised method learns what normal looks like and flags points that differ.
SageMaker Random Cut Forest (RCF)
• RCF builds many random trees, each by repeatedly cutting the data along randomly chosen dimensions.
• Normal points sit in dense regions and need many cuts to isolate; unusual points are isolated after few cuts.
• Each data point receives an anomaly score; a common rule flags scores more than about three standard deviations above the mean score.
• Training needs no labels. You can use shingling (sliding windows) so the model sees short sequences of a metric instead of single values.
Example: CPU utilization normally ranges from 20% to 60%; a sudden jump to 98% at 3 a.m. receives a high score and triggers an alert.
Why each option works or fails
• ✗ A. XGBoost classifier: a classifier learns from labeled examples of each class, and there is no labeled anomaly data to train it on.
• ✓ B. Random Cut Forest (correct): it is an unsupervised algorithm that assigns anomaly scores to data points, so it detects spikes without any labeled anomalies.
• ✗ C. Factorization Machines: this is a supervised algorithm for sparse, high-dimensional data such as recommendations; it needs labels and is not an anomaly detector.
• ✗ D. Linear Learner: Linear Learner performs supervised regression or classification, which requires target labels that do not exist here.
🎯 Key idea: No labels plus the need to find unusual points means unsupervised Random Cut Forest; supervised algorithms need labeled anomalies.
🔗 More from this exam
• All MLA-C01 practice sets
• Every exam on this blog
📚 Keep studying MLA-C01
More MLA-C01 questions: https://hyunhp.github.io/cert-bank/#exam=MLA-C01
Study guide: https://hyunhp.github.io/cert-bank/study.html#exam=MLA-C01
These are original practice questions, not real exam questions, and cert-bank is not affiliated with any certification vendor.
Original practice questions written for this blog, not exam dumps.
Comments
Post a Comment