CAIP (AIP-210) study plan

Define supervised, unsupervised, self-supervised, and reinforcement learning and classify ten example tasks into the correct paradigm with a one-line justification each.

Distinguish classification from regression, and precision, recall, F1, ROC-AUC, and MAE/RMSE, then state which metric you would report for five described business scenarios.

Explain the bias-variance trade-off and overfitting/underfitting in your own words and identify which of six described training curves shows each condition.

Describe how a train/validation/test split and k-fold cross-validation differ and when each is appropriate for small versus large datasets.

Compare at least six algorithm families (linear and logistic regression, decision trees, random forests, gradient boosting, k-NN, k-means, neural networks) on interpretability, data needs, and typical use cases.

Explain how a neural network trains: forward pass, loss function, backpropagation, gradient descent, and the role of learning rate and epochs.

Distinguish bagging from boosting and explain why gradient-boosted trees often outperform single decision trees on tabular data.

Describe dimensionality reduction and clustering use cases and when to prefer PCA versus t-SNE for visualization.

Explain the transformer architecture at a conceptual level: self-attention, positional encoding, encoder-decoder structure, and why it replaced recurrent models for many NLP tasks.

Distinguish pretraining from fine-tuning, and describe what embeddings represent and how vector similarity search is used in retrieval.

Compare foundation model, fine-tuned model, and RAG approaches on cost, latency, data freshness, and accuracy for three described requirements.

Describe how generative models produce text and images and name two evaluation approaches for generative output quality.

Map the AWS AI/ML service landscape: SageMaker components (Studio, Training, Processing, Endpoints, Pipelines, Model Registry), plus Bedrock, Rekognition, Comprehend, Transcribe, and Forecast, and match each to a described use case.

Explain the difference between batch transform and real-time inference endpoints and choose one for four described workloads.

Translate a stated business problem into a formal ML problem statement with target variable, features, success metric, and baseline.

Identify data quality risks (leakage, imbalance, missing values, drift) in a described dataset and propose one mitigation for each.

Design an end-to-end ML pipeline covering data ingestion, preprocessing, feature engineering, training, evaluation, and deployment, and justify each stage.

Select appropriate feature engineering and preprocessing steps (encoding, scaling, imputation, resampling) for three described datasets.

Choose hyperparameters and a tuning strategy (grid, random, Bayesian) for a described model and justify the compute trade-off.

Explain how to detect and mitigate class imbalance and describe the effect on the chosen evaluation metric.

Compare deployment patterns (real-time endpoint, serverless inference, batch, asynchronous) and select one for four workloads with cost and latency justification.

Describe how to version data, code, and models, and explain the role of a model registry in reproducibility and rollback.

Explain monitoring requirements after deployment: data drift, model drift, latency, error rates, and how automated retraining is triggered.

Describe A/B testing and shadow deployment for model rollout and state the metric you would compare.

Identify security and compliance controls for an AI solution: IAM least privilege, encryption at rest and in transit, VPC isolation, and data residency considerations.

Explain cost drivers of ML workloads (training compute, endpoint uptime, storage, inference volume) and propose three concrete cost optimizations.

Describe responsible AI practices: bias detection, explainability, human review, and model documentation, and apply them to a described high-stakes use case.

Explain the handoff package an ML engineer delivers to operations: documentation, runbooks, access procedures, and acceptance criteria.

Re-take and review every self-quiz from weeks 1 through 7, logging each missed concept and restudying it until you can explain it without notes.

Practice scenario-style questions that combine domains, such as choosing a service, metric, and deployment pattern for a single described business case.

Time yourself on a 40-question mixed practice set and analyze error patterns by domain to target final review.

Write a one-page summary of the ten concepts you are least confident about and how each would appear in an exam scenario.