Fundamentals of AI and ML — AWS Certified AI Practitioner AIF-C01 practice questions

Which BEST describes deep learning?

1. A rules-based expert system with hand-coded logic
2. A relational schema optimized for analytical queries
3. ML using neural networks with many hidden layers
4. A statistical technique for linear regression
5. A cloud service for hosting containerized workloads

Answer: C — ML using neural networks with many hidden layers

A) Symbolic AI. B) Data warehouse. C) Correct — DL is ML with deep neural networks (vision, NLP, speech, GenAI). D) Classical stats. E) Container platform.

Which set of metrics is BEST for a binary classification model?

1. BLEU, ROUGE, METEOR, chrF, and BERTScore
2. MSE, MAE, RMSE, and coefficient of determination
3. Latency, throughput, memory, and instance cost
4. Silhouette, Davies-Bouldin, and Calinski-Harabasz
5. Precision, recall, F1, accuracy, and AUC-ROC

Answer: E — Precision, recall, F1, accuracy, and AUC-ROC

A) NLG evaluation. B) Regression metrics. C) Serving metrics. D) Clustering metrics. E) Correct — standard classification metrics.

How do training and inference costs typically compare?

1. Training is one-time and expensive; inference is ongoing and per-call
2. Inference is always more expensive than training in total
3. Both are billed identically per hour regardless of workload
4. Training is free on AWS; inference is billed per request
5. Inference is charged only when the model produces errors

Answer: A — Training is one-time and expensive; inference is ongoing and per-call

A) Correct — training is a large lump-sum; inference cost accrues per token/request over time. B) Only true at very high steady inference volumes vs. small models. C) False. D) False. E) False.

Which AWS service is BEST for end-to-end ML development, training, and deployment?

1. Amazon Bedrock model API
2. Amazon SageMaker platform
3. Amazon Comprehend NLP
4. AWS Glue ETL service
5. Amazon Kendra search

Answer: B — Amazon SageMaker platform

A) Serves FMs, not general ML lifecycle. B) Correct — full lifecycle: notebooks, training, tuning, hosting, MLOps. C) NLP only. D) ETL. E) Search.

Which statement about deep learning is CORRECT?

1. Deep learning is unrelated to machine learning
2. Deep learning is a synonym for symbolic AI
3. Deep learning is ML using neural networks with many layers
4. Deep learning is limited exclusively to image workloads

Answer: C — Deep learning is ML using neural networks with many layers

A) DL is a subset of ML. B) Symbolic AI is rules-based, not DL. C) Correct — DL powers vision, NLP, speech, and GenAI. D) DL spans many modalities.

Which contrast between supervised and unsupervised learning is CORRECT?

1. Supervised uses labeled data; unsupervised finds patterns without labels
2. Supervised is unsupervised at inference; unsupervised is labeled at training
3. Supervised only handles regression; unsupervised only handles classification
4. Supervised requires GPUs; unsupervised runs on CPUs only
5. Supervised uses embeddings; unsupervised uses one-hot vectors

Answer: A — Supervised uses labeled data; unsupervised finds patterns without labels

A) Correct — the defining distinction. B) Backwards nonsense. C) Reversed and false. D) Hardware is not the distinction. E) Not a definitional split.

On a highly imbalanced dataset (99% negatives), which is the WEAKEST single metric to optimize?

1. Overall classification accuracy
2. Precision-recall AUC for the minority class
3. F1 score computed for the positive class
4. Recall on the positive (minority) class

Answer: A — Overall classification accuracy

A) Correct — 'always predict negative' scores 99% accuracy but is useless. B) PR-AUC handles imbalance. C) F1 targets both precision and recall. D) Recall directly measures caught positives.

Which BEST describes feature engineering?

1. Choosing an AWS region that minimizes training latency
2. Creating and selecting features that improve model performance
3. Selecting which GPU instance family to use for training
4. Deciding how many CloudWatch alarms to attach to the endpoint

Answer: B — Creating and selecting features that improve model performance

A) Region planning is unrelated. B) Correct — one-hot encoding, scaling, aggregations, interactions, embeddings. C) Compute choice. D) Ops monitoring.

Which BEST describes the AIF-C01 exam structure?

1. 100 questions, 180 minutes, oral defense, $600 fee
2. 40 questions, 30 minutes, passing at 900/1000, $30 fee
3. 250 questions, 6 hours, adaptive testlets, $1200 fee
4. 65 questions, 90 minutes, passing at 700/1000, $100 fee
5. 200 questions, 4 hours, essay grading, $500 fee

Answer: D — 65 questions, 90 minutes, passing at 700/1000, $100 fee

A/B/C/E) Wrong numbers for AIF-C01. D) Correct — foundational cert, no prerequisites.

Which BEST describes machine learning?

1. A subfield of AI where systems learn patterns from data
2. A rules engine that runs explicitly programmed logic
3. A relational database used to query structured records
4. A dashboarding tool that visualizes trends in metrics
5. A hardware accelerator designed for tensor operations

Answer: A — A subfield of AI where systems learn patterns from data

A) Correct — ML learns from examples rather than fixed rules. B) Describes symbolic AI / expert systems. C) Describes an RDBMS. D) Describes BI tooling. E) Describes GPUs/TPUs.

SageMaker Pipelines?

ML workflow orchestration. Define steps as DAG. CI/CD for ML.

Supervised vs unsupervised learning?

Supervised: labeled data. Unsupervised: unlabeled, finds patterns.

AIF-C01 exam approach?

Concepts over hands-on. Know AWS service positioning, responsible AI, basic ML terms.

Validation vs test data?

Validation: tune hyperparameters during training. Test: final evaluation, untouched until end.