🌟 The Ultimate Memory Hooks for AWS Certified AI Practitioner (AIF-C01)

Source: Dev.to

Machine Learning Basics

Supervised vs Unsupervised

• Labels → Supervised
• No Labels → Unsupervised

✔ Supervised = Teacher + Correct Answers
✔ Unsupervised = Find patterns (clustering, segments)

Classification vs Regression

• Classes → Classification
• Numbers → Regression

Overfitting vs Underfitting

• Overfitting = Too complex → Increase regularization
• Underfitting = Too simple → Decrease regularization

Key Algorithms

• Clustering – Group customers? No labels? → K‑Means
• Image Classification – Flower classification → k‑NN or Decision Tree
• Anomaly Detection – No labels + abnormal detection → Autoencoders

GenAI Prompt Engineering

• Few‑shot prompting – Show format → Few‑shot prompting.
• Prompt chaining – Multi‑step workflow → Prompt chaining.
• ReAct prompting – Reason + Action + Tool use → ReAct.

Temperature

• Creativity ↑ → Temperature ↑
• Consistency ↑ → Temperature ↓

LLM Inference Parameters

• Temperature – Creativity
• Top‑K – Number of token choices
• Top‑P – Probability bucket
• Max Tokens – Output length
• Frequency Penalty – Reduce repeated words
• Presence Penalty – Discourage repeated topics

Mapping

• Creativity → Temperature / Top‑K / Top‑P
• Length → Max Tokens
• Repetition → Frequency & Presence

Retrieval‑Augmented Generation (RAG)

Purpose of Chunking

Chunking = Better retrieval → Better context

Batch Steps in RAG

• ✔ Content embeddings
• ✔ Build search index

Do not include query embeddings or response generation in this batch.

LLM Type for Multimodal Search

Text + Image queries → Multimodal model

Evaluating ML Models

Summarization Metrics

• ROUGE (default)
• If ROUGE missing → Choose BLEU

Translation Metrics

• BLEU / METEOR

Classification Metrics

• Imbalanced data → F1 Score
• Balanced data → Accuracy

Regression Metrics

• Numeric prediction → MSE / RMSE

LLM Quality

• Perplexity – How surprised is the model?

AWS Services – Quick Memory Hooks

• Model Cards – Governance + Documentation
• Model Monitor – Detect drift in production
• Ground Truth – Human labeling
• JumpStart – Pre‑built models + quick deploy
• SageMaker Canvas – No‑code data prep
• HealthScribe – Medical speech‑to‑text
• Guardrails for Bedrock – Responsible AI (safety filters)
• PartyRock – Experiment + Learn + No cost (Not for VPC, not for deployments)

GenAI Lifecycle

Design → Data → Train → Evaluate → Deploy → Monitor

Evaluation Stage

• Accuracy testing
• Safety + toxicity testing
• Hallucination measurements

Inference

• Train = Learn
• Infer = Predict
• Deploy = Serve

Embeddings

• Embeddings = Meaning → Vectors
• Reduced dimension → Same meaning → Similarity search

Foundational Concepts

Fine‑tuning

Teach a large model a small task well.

• Domain‑specific labeled data
• Improves specific task performance
• Not retraining from scratch
• Not updating model to recent events

Responsible AI

Safety + Filters + Detect toxicity → Use Guardrails

Final Thoughts

These memory hooks are designed to:

• Make recall instant during the exam
• Reduce confusion between similar concepts
• Build confidence with patterns instead of memorising definitions

Prepared using insights from the QA/CloudAcademy course “AWS Certified AI Practitioner (AIF‑C01) Certification Preparation” by Danny Jessee.