1.4. Reflection Checkpoint: Fundamentals of AI and ML

Key Takeaways

• AI, ML, and Deep Learning nest inside each other (AI ⊃ ML ⊃ DL) — DL is what powers modern LLMs, but not every ML problem needs deep neural networks.
• The three learning paradigms are distinguished by what the data looks like: supervised needs labeled data, unsupervised works with unlabeled data to find structure, and reinforcement learning learns from reward/penalty feedback in an environment.
• Precise vocabulary matters: an algorithm trains a model, and a model is used for inferencing — conflating these terms creates confusion on the exam and in practice.
• Whether a business question calls for regression ("how much"), classification ("which category"), or clustering ("what groups") should be the first question you ask before picking a technique — and AI/ML itself is only appropriate for prediction/pattern problems, not deterministic calculations.
• AWS managed AI services (Rekognition, Comprehend, Textract, Transcribe, Polly, Lex, Kendra, Forecast, Personalize, Fraud Detector) solve common problems via API call — reach for these before building a custom SageMaker model.
• The ML lifecycle (problem formulation → data collection → EDA → preprocessing → training → evaluation → tuning → deployment → monitoring) is iterative, not linear — monitoring can trigger a full retrain.

Connecting Forward

Phase 1 built your foundation in traditional ML: how models learn, how AWS supports the lifecycle, and how to match a business problem to the right technique. Phase 2 narrows in on a specific, heavily-tested category of ML that works differently — Generative AI, including foundation models, and the AWS infrastructure that supports them.

Self-Check Questions

• A team explains they're "running the model" to identify fraud patterns in a dataset with no labeled fraud/not-fraud examples. Is this really "running a model," and which learning paradigm actually applies here?
• Your manager asks for an AI system that flags any transaction over $10,000 for manual review — a fixed, deterministic rule. Should this be built as an ML model? Why or why not?