Amazon AIF-C01: Fundamentals of AI and ML — Study Guide
Part of the AWS AI Practitioner AIF-C01 — Study Guide. Practice with verified answers in the Amazon exam hub, or take timed practice tests on ExamRoll.io.
Core Concepts and Model Types
Understanding core terminology anchors sound architecture and trade‑off decisions. Supervised learning maps inputs to labeled outputs and splits into regression for continuous targets and classification for discrete classes; multiclass classification (20 gene classes) versus binary decisions changes model choice and loss functions. Unsupervised learning finds structure in unlabeled data through clustering and dimensionality reduction for segmentation and anomaly detection. Common algorithm families include linear models and tree ensembles (interpretable, fast for tabular data), kernel methods (SVMs), and neural networks (flexible, high-capacity for images, audio, and text). For image tasks, convolutional networks or pretrained vision foundation models (via Amazon SageMaker JumpStart or Amazon Rekognition Custom Labels) dominate; for text, transformer architectures power large language and sequence models. Interpretability requirements push practitioners toward simpler models or explainability tools such as SageMaker Clarify. Model selection balances predictive performance, interpretability, inference latency, and cost: a tree-based model may suffice for gene classification if transparency is required, whereas deep models fit high-dimensional image or language tasks. Knowing these families and when to reuse pretrained models versus training from scratch guides efficient, compliant solutions.
ML Lifecycle, Data and Feature Management
The machine learning lifecycle begins with problem framing, data collection, labeling, feature engineering, training, validation, deployment, monitoring, and iteration. In AWS environments, SageMaker orchestrates many lifecycle stages: SageMaker Processing for data transforms, SageMaker Feature Store for centralized feature storage and online/offline serving, and SageMaker Pipelines for CI/CD of model workflows. Label quality and class balance are frequent bottlenecks; invest in robust annotation pipelines or use SageMaker Ground Truth to reduce noisy labels and active learning to focus human effort. Feature lineage and access controls matter for reproducibility and governance; register features and models in SageMaker Model Registry and version datasets in Amazon S3 with lifecycle policies. For production, include automated model monitoring with SageMaker Model Monitor to detect data drift, concept drift, and accuracy degradation, and integrate alerts via Amazon CloudWatch and AWS Lambda. Architect pipelines with idempotency and recovery: use event-driven patterns (S3 events, Step Functions) and design compute and storage to scale—training on EC2/GPU or Trainium, inference on Inf1/Neptune/Graviton instances—based on latency and throughput requirements.
Evaluation, Algorithms, and Common Traps
Selecting evaluation metrics and baselines is a decisive practice. For classification problems consider precision, recall, F1-score, and area under the ROC curve; for multiclass use per-class recall and macro/micro averages. Regression evaluation uses RMSE, MAE, and R-squared. For production serving, runtime efficiency is measured by tail latency (P95/P99) and throughput (requests per second), and cost per inference. Typical traps include choosing accuracy in imbalanced datasets, overfitting through excessive hyperparameter tuning on test data, and neglecting calibration when probabilities are used for business decisions. Use cross-validation and holdout validation, and implement baseline models (simple heuristics) to ensure ML adds value. When labeled data is scarce, use transfer learning, data augmentation for images, or semi-supervised approaches. For explainability and compliance, pair opaque models with post-hoc explainers (SHAP, integrated gradients) and integrate SageMaker Clarify. Common algorithm choices on AWS: XGBoost for tabular with fast SageMaker training, CNNs for images via Torch/TensorFlow on GPU instances, and transformers for text using Hugging Face on SageMaker or Bedrock foundation models for retrieval-augmented generation.
Foundation Models, Deployment Patterns, and Security
Foundation models accelerate development but introduce distinct architectural choices and security considerations. Amazon Bedrock provides managed access to multiple base models and supports fine‑tuning, retrieval‑augmented generation (RAG) workflows, and integration with vector stores such as Amazon OpenSearch Service (k-NN) or Amazon Kendra for semantic search. Choose between fine‑tuning, instruction‑tuning, or lightweight adapters based on data size and safety needs. For low-latency, high-throughput inference consider deploying optimized models on Amazon EC2 Inf1 (Inferentia) or using SageMaker Neo/Edge Manager to compile and host models on edge devices; absolute lowest latency on-device inference requires model quantization, pruning, and local runtime via SageMaker Edge Manager or AWS IoT Greengrass, trading off model size and accuracy. Security patterns include VPC endpoints, AWS PrivateLink for Bedrock, strict IAM roles, KMS‑backed S3 encryption for sensitive data, and audit logging. To reduce hallucination and prompt attacks, implement guardrails: input sanitization, prompt templates, response filters, and RLHF or post‑processing checks. Monitor usage and anomalous API calls with CloudTrail and implement rate limits and anomaly detection to protect model access and data privacy.
Practical Problem: Use-Case Scenario
Scenario: GreenGene Labs uses an AWS AI environment with Amazon S3 for raw data, SageMaker for model development, and Amazon Bedrock for experimenting with foundation models. They maintain sensitive genomic and clinical metadata under strict access controls and need scalable inference for research dashboards.
Challenge: Classify human gene samples into 20 categories with transparent decision logic, maintain feature reuse across teams, and deploy a monitored, low-latency API for researchers.
Recommended Approach:
- Use SageMaker Processing and Ground Truth to prepare and label data, storing curated features in SageMaker Feature Store.
- Train interpretable models (XGBoost, decision trees) in SageMaker Training; register models in SageMaker Model Registry and record lineage.
- Deploy the best model to SageMaker endpoint behind an Application Load Balancer with autoscaling; enable SageMaker Model Monitor for drift and CloudWatch alerts.
- For experimentation with larger representations, use Bedrock or Hugging Face models for embeddings and add an OpenSearch vector index for similarity search; combine with the interpretable model for final classification.
Rationale: Centralized feature management and reproducible pipelines reduce data leakage and speed collaboration, while prioritizing interpretable models satisfies transparency needs and Bedrock embeddings enable richer representation without sacrificing governance.
All domains · Generative AI Concepts →
Practice these questions → · Timed practice on ExamRoll.io →
Pass the whole exam — not just this question
You found this answer. Get every verified question and explanation in one place, and save hours of prep. Free to start.
Pass your exam →