AWS AIF-C01 — Day 2 Comprehensive Reference Guide

Before any AWS service or specific algorithm, you need a clean mental map of what machine learning is, what kinds of problems it solves, and how a project flows from raw data to a deployed model. This section gives you that map.

Every exam scenario is a walk down this four-step ladder:

Master this ladder and the exam stops being a memorization contest. Every story — a hospital detecting tumors, a streaming service suggesting movies, a factory finding defects — is a guided walk down those four rungs.

2.1 Supervised Learning

The intuition: A child learning what a cat is. You show photos and say "this is a cat" or "this is not a cat." Each example has the right answer attached.

Mechanics:

• Input data is labeled (every example has features + correct answer)
• Model learns the mapping: features → label
• Two flavors: classification (label is a category) and regression (label is a number)

Real examples:

• Loan approval: features (age, salary, credit score) → label (approved / rejected)
• Tumor diagnosis: image → label (malignant / benign)
• Demand forecast: historical sales → label (next week's sales number)

2.2 Unsupervised Learning

The intuition: Same child, but now you dump a pile of photos with no labels and say "sort these however you want." The child finds structure that was already there but hidden.

Mechanics:

• Input data has no labels
• Model finds hidden structure: clusters, patterns, outliers
• You don't tell it what's "correct" — it discovers it

Real examples:

• Customer segmentation: group 10 million customers into "buying personas" you didn't define
• Topic discovery: find themes in a million news articles
• Anomaly detection: find the 0.01% of credit card transactions that "look weird"

2.3 Reinforcement Learning (RL)

The intuition: Training a dog. The dog tries something → you give a treat (reward) or say "no" (penalty). Over many trials, the dog learns which actions lead to treats.

Mechanics:

• An agent takes actions in an environment
• Each action returns a reward (positive or negative)
• The agent's goal: learn a policy (strategy) that maximizes total reward over time
• Learning happens via trial and error

Real examples:

• Game-playing AI (AlphaGo, chess engines)
• Robotics (a robot learning to walk)
• Autonomous driving simulation
• Dynamic pricing (raise/lower price → observe sales → adjust)
• Ad bidding strategies

2.4 Semi-supervised Learning

The intuition: You have 1,000 medical X-rays carefully labeled by a radiologist (very expensive!) and 100,000 unlabeled X-rays sitting in a database. Throwing away the unlabeled ones is wasteful.

Mechanics:

• Small amount of labeled data + large amount of unlabeled data
• Model uses unlabeled data to learn structure, labeled data to anchor meaning
• Best of both worlds when labeling is expensive

Real examples:

• Medical imaging (labeling requires doctors)
• Speech recognition (transcription is slow)
• Web page classification (millions of pages, few labeled)

🧠 Quick check — try these in your head:

The "problem type" answers: What shape is the output?

This is different from "type of ML." A supervised problem can be classification or regression — those are different problem types.

3.1 Classification — predict a category

Output is discrete (a finite set of choices).

• Binary classification: 2 classes (spam vs not spam)
• Multi-class classification: 3+ classes (cat / dog / bird)
• Multi-label classification: multiple labels per input (an image can be "beach" AND "sunset" AND "people")

3.2 Regression — predict a number

Output is continuous (a number on a scale).

Key distinction from classification: "Will this customer churn?" = classification. "How many days until they churn?" = regression.

3.3 Clustering — group similar things (unsupervised)

No labels exist. Model finds groups.

Examples:

• Customer segmentation (find 5 natural buyer personas)
• Document grouping (organize 1M articles into themes)
• Genome analysis (group similar gene expression profiles)

3.4 Anomaly Detection — find the weird stuff

Find rare events that don't match the normal pattern.

Examples:

• Unusual login (different country, weird time)
• Fraudulent transaction (purchase pattern doesn't match user)
• Machine failure (vibration sensor showing abnormal pattern)
• Network intrusion (traffic pattern looks unlike usual)

Why it's special: You usually have very few anomaly examples. So pure supervised learning struggles. Often handled as unsupervised or one-class.

3.5 Recommendation — suggest relevant items

Predict what a user will like based on their history and similar users' behavior.

Examples:

• Netflix: "because you watched X..."
• Amazon: "customers who bought this also bought..."
• Spotify Discover Weekly
• News article recommendations

🧠 Map the problem to the type

Every real ML system moves through these 7 stages. Each AWS feature you'll memorize maps to one of them.

🧠 Quick check — Walk through one full scenario:

That's the 7-stage loop, end to end.

Once you understand what ML is, the next layer is the materials and machinery. Data, parameters vs hyperparameters, and training vs inference. These distinctions create most of the exam traps.

5.1 Labeling

Adding the correct answer to each training example. Required for supervised learning.

• Done by humans, ML, or both
• AWS service: SageMaker Ground Truth (uses active learning to auto-label easy cases as model improves)

5.2 Data Splits — the cardinal rule

Never use all data for training. Split it:

5.3 Class Imbalance

When one class dominates (fraud is 1% of transactions, rare disease is 0.1% of patients).

Why accuracy lies here: A model that always predicts "legitimate" is 99% accurate but useless.

Fixes:

5.4 Data Augmentation

Artificially expand your training set by creating modified versions of existing examples.

Why it works: More effective data + teaches model to be robust to variations.

This distinction is small but heavily tested.

6.1 Parameters

What: Numbers the model learns automatically during training.

Examples:

• Weights in a neural network
• Biases
• Learned word embeddings
• Split points in a decision tree

You never set these by hand. They emerge from training.

6.2 Hyperparameters

What: Numbers you choose before training.

Examples:

• Learning rate
• Batch size
• Number of epochs
• Number of layers / neurons (for NN)
• Number of trees / max depth (for trees)
• Regularization strength

You don't know the best values in advance — you search.

6.3 The clean distinction

6.4 Hyperparameter Tuning

You search for good hyperparameters. AWS: SageMaker Automatic Model Tuning (uses Bayesian optimization).

🧠 Quick check:

Two phases in the life of a model. Different characteristics, different costs.

7.1 Inference Deployment Types (4 ways to serve a model)

A model that runs is not a model that works. This section teaches you to recognize subtle failures, describe them with bias/variance language, and pick the right metric.

The single most important concept in ML.

8.1 Overfitting

The intuition: A student who memorized every practice exam answer. Aces practice exams. Fails the real one. They never learned — they memorized.

Symptom:

Big gap = overfitting.

Causes:

• Model too complex (too many parameters)
• Not enough training data
• Trained for too many epochs
• Too many noisy features

Fixes:

8.2 Underfitting

The intuition: A student who only studied the chapter titles. Fails practice exams. Fails the real one. Never learned anything substantive.

Symptom:

Both low, small gap = underfitting.

Causes: Too-simple model, too few features, too little training, wrong algorithm.

Fixes: More complex model, more features, train longer, less regularization.

🧠 Quick diagnosis check:

Mathematical language for the same overfit/underfit story.

🧠 Quick check — three teams predicting house prices:

Team C wins.

The single most heavily tested area. Wrong metric = wrong answer.

10.1 The Confusion Matrix (foundation of all classification metrics)

• TP: Said positive, was positive ✅
• FP: Said positive, was negative (false alarm) ⚠️
• FN: Said negative, was positive (missed a real one) ⚠️
• TN: Said negative, was negative ✅

10.2 Accuracy

Accuracy = (TP + TN) / Total

Plain English: Of all predictions, how many were right?

Use when: Classes are roughly balanced.

10.3 Precision

Precision = TP / (TP + FP)

Plain English: Of everything I flagged as positive, how many actually were?

Use when: False positives are expensive. You want to be sure when you say "positive."

Classic example — Spam filter:

• Marking a real email as spam (FP) = user misses something important
• High precision = "When I say spam, it's really spam"

10.4 Recall (Sensitivity)

Recall = TP / (TP + FN)

Plain English: Of all the actual positives out there, how many did I catch?

Use when: False negatives are expensive. Can't afford to miss a positive case.

Classic example — Cancer screening:

• Missing a real cancer case (FN) = patient doesn't get treated
• High recall = "I catch nearly every real cancer case, even with some false alarms"

10.5 The Precision–Recall Tradeoff

You trade one for the other by adjusting the decision threshold.

• Lower threshold → catch more (↑ recall) but more false alarms (↓ precision)
• Raise threshold → fewer false alarms (↑ precision) but miss some (↓ recall)

Choose based on what's costlier in your domain.

10.6 F1 Score

F1 = 2 × (Precision × Recall) / (Precision + Recall)

Plain English: Harmonic mean. Single number balancing both. Bad if either is bad.

Use when:

• Imbalanced data
• Both FP and FN matter
• You want one combined metric

10.7 AUC-ROC

ROC curve: Plots True Positive Rate vs False Positive Rate at every threshold. AUC: Area Under that Curve, a single number 0–1.

Use when: Comparing classifiers without committing to a threshold.

10.8 Regression Metrics

🧠 Metric scenario quick-fire:

The engines of modern AI. Tested at conceptual level — understand intuitions, no math required.

11.1 The Neuron

Each neuron:

1. Takes numerical inputs
2. Multiplies each by a weight
3. Adds them up + a bias
4. Runs result through an activation function
5. Outputs a number

Weights and biases = parameters (learned during training).

11.2 Layers

Deep learning = many hidden layers.

11.3 Activation Functions

11.4 Backpropagation (how it learns)

1. Forward pass: Input → output → prediction
2. Compute error (prediction vs correct answer)
3. Backward pass: Error flows backward, layer by layer
4. Update weights in the direction that reduces error
5. Repeat for every batch, thousands of times

Three architectures dominate three kinds of data. Memorize this.

12.1 CNN — Convolutional Neural Network

Best for: images and visual data.

Why: Uses small filters that slide over the image to detect local patterns (edges → textures → shapes → objects).

Use cases:

• Image classification (cat / dog / car)
• Object detection
• Facial recognition
• Medical imaging
• Self-driving car perception
• Manufacturing defect detection

12.2 RNN — Recurrent Neural Network

Best for: sequential data where order matters.

Why: Has memory — output at each step depends on current input + hidden state from previous step.

Use cases:

• Time series forecasting
• Sensor data analysis
• (Older) language and speech models

Weakness: Struggles with long-range dependencies — forgets the start of a long sequence.

Modern variants: LSTM, GRU (better memory).

12.3 Transformers

Best for: language and generative AI. The architecture behind every modern LLM (GPT, Claude, Llama).

Why they dominate: Instead of step-by-step like RNNs, transformers use an attention mechanism that lets every token directly look at every other token.

• Handles very long sequences
• Trains in parallel (massively faster)
• Scales to billions of parameters

Use cases:

• Text generation
• Translation
• Summarization
• Question answering
• Chatbots
• Code generation
• All Large Language Models / Foundation Models

🧠 Quick mapping:

A model in a notebook is not a useful model. MLOps is the discipline of getting them into production safely.

MLOps = DevOps applied to ML.

13.1 Model Versioning

Track every version of every model.

Why:

• Rollback if a new model breaks production
• Compare v2 vs v3 performance
• Audit what changed and when
• Reproduce past results
• Compliance (GDPR, banking regulations)

13.2 CI / CD / CT for ML

13.3 Production Monitoring

13.4 Retraining Triggers

You retrain when:

• Accuracy drops below threshold
• New data becomes available
• Data drift detected
• Business rules change (new product launch, regulation)
• Seasonality changes
• Bias detected
• Compliance requires it

AWS service: SageMaker Pipelines automates this loop end to end.

🧠 Quick scenario:

That's the MLOps loop in action — possible only because the tooling was already in place.

The AWS flagship ML platform. The single most heavily tested service on the exam.

The exam tests which feature handles which stage of the ML lifecycle. Memorize these.

14.1 SageMaker Studio

• What it is: Web-based IDE for ML development
• Use when: You want one workspace to build, train, debug, deploy, monitor
• 🔑 Clue: "integrated development environment for ML"

14.2 SageMaker Autopilot

• What it is: AutoML — give it data, get a deployable model
• Use when: Minimal coding / minimal ML expertise needed
• 🔑 Clue: "AutoML", "automatically build model", "least amount of code"

14.3 SageMaker Ground Truth

• What it is: Data labeling service (humans + active learning)
• Use when: You need labeled training data for supervised learning
• 🔑 Clue: "label training data", "human labelers"

14.4 SageMaker Pipelines

• What it is: Workflow orchestration for ML (preprocess → train → evaluate → deploy)
• Use when: Repeatable MLOps pipelines / CI-CD for ML
• 🔑 Clue: "orchestrate ML workflow", "automate end-to-end ML"

14.5 SageMaker Clarify

• What it is: Detects bias and explains predictions (feature attribution)
• Use when: Fairness checks or explainability needed
• 🔑 Clue: "bias detection", "explainability", "why did the model predict this?"

14.6 SageMaker Model Monitor

• What it is: Watches deployed models in production for drift
• Use when: Detect data drift, model drift, prediction quality decline
• 🔑 Clue: "monitor production model", "data drift", "model drift"

14.7 SageMaker JumpStart

• What it is: Catalog of pretrained models, foundation models, solution templates
• Use when: You want a quick start without training from scratch
• 🔑 Clue: "pretrained model", "foundation model in SageMaker", "solution template"

14.8 SageMaker Feature Store

• What it is: Central repository for ML features, used in training AND inference
• Use when: Multiple teams reusing features, training/inference consistency needed
• 🔑 Clue: "reuse features", "feature consistency", "central store for features"

14.9 SageMaker Data Wrangler

• What it is: Visual tool for data prep and feature engineering
• Use when: Want low-code data cleaning/transformation

14.10 SageMaker Canvas

• What it is: No-code ML model building for business analysts
• Use when: Non-developers need to build models

🧠 Scenario → SageMaker feature:

Three options, trade off control vs effort.

15.1 Common Built-in Algorithms

You're not building a model — you're calling a service that already has one. Faster, cheaper, no ML expertise needed.

Both deal with visual input but answer different questions.

17.1 Amazon Rekognition

What: Image and video analysis.

Capabilities:

• Detect objects, faces, scenes, labels
• Moderate unsafe / inappropriate content
• Recognize celebrities
• Detect text in images (signs, captions)
• Facial analysis (age, emotion, attributes)
• Compare faces

17.2 Amazon Textract

What: Document data extraction (OCR + structure).

Capabilities:

• OCR (image of text → text)
• Extract tables (rows, columns)
• Extract forms (field labels + values)
• Extract key-value pairs from invoices, receipts
• Handle PDFs, scanned documents

🧠 Rekognition vs Textract — the trap:

18.1 Amazon Comprehend

What: Natural language processing on text.

Capabilities:

• Sentiment analysis (positive / negative / neutral / mixed)
• Entity extraction (people, places, organizations, dates)
• Key phrase extraction
• Language detection
• Topic modeling on document collections
• Custom classification
• Comprehend Medical (medical entities, drugs, conditions)

18.2 Amazon Translate

What: Language translation (real-time + batch).

18.3 Amazon Transcribe

What: Speech to text.

Capabilities:

• Real-time and batch
• Multiple speakers (diarization)
• Custom vocabulary
• Many languages
• Transcribe Medical for clinical dictation

18.4 Amazon Polly

What: Text to speech (the opposite of Transcribe).

Capabilities:

• Many languages and voices
• Neural voices (very natural)
• Used for accessibility, voice assistants, audiobooks

🧠 Transcribe vs Polly — the mirror:

19.1 Amazon Lex

What: Chatbot service (same engine as Alexa).

Two key concepts:

• Intents — what the user wants to do (BookFlight, CheckBalance)
• Slots — info needed to fulfill the intent (departure city, date, passengers)

19.2 Amazon Kendra

What: Intelligent enterprise search (NL question answering across documents).

Capabilities:

• Natural language questions, not just keywords
• Connects to S3, SharePoint, Confluence, Salesforce, etc.
• Returns actual answers, not just doc lists

🧠 Lex vs Kendra:

20.1 Amazon Forecast

What: Time-series forecasting service.

Use cases: Demand forecasting, sales forecasting, inventory planning, workforce demand.

20.2 Amazon Personalize

What: Recommendation engine (same tech as amazon.com).

Use cases: Product recommendations, content recommendations, personalized rankings, similar items.

🧠 Forecast vs Personalize — the trap:

20.3 Amazon Fraud Detector

What: Dedicated fraud detection (online payments, fake accounts, transaction risk).

20.4 Amazon Lookout for Metrics

What: Detect anomalies in business/operational metrics (revenue, traffic, conversion rates).

20.5 Other Lookout Services

21.1 Amazon Bedrock

What: Fully managed service for foundation models (Claude, Llama, Titan, etc.).

Capabilities:

• Single API to many models
• No infrastructure to manage
• Sub-features:
• Knowledge Bases — RAG with your own documents
• Agents — multi-step workflows that call APIs
• Guardrails — safety policies on prompts and outputs
• Model Evaluation — compare models on your benchmarks

21.2 Amazon Q

What: Generative AI assistant (finished product, built on foundation models).

🧠 Bedrock vs JumpStart vs Q:

Pre-exam revision. Read it three times the night before.

• Supervised = labeled data
• Unsupervised = no labels, find patterns
• Reinforcement = agent + reward
• Semi-supervised = few labels + many unlabeled
• Classification = predict category
• Regression = predict number
• Clustering = group similar
• Anomaly detection = find unusual
• Recommendation = suggest items
• Overfitting = memorized training (huge train/test gap)
• Underfitting = too simple (both low)
• Bias = bad assumptions; Variance = volatile output
• Precision = avoid false positives (spam filter)
• Recall = avoid false negatives (cancer screening)
• F1 = balance both (imbalanced data default)
• AUC-ROC = class separation across thresholds
• CNN = images; RNN = sequences; Transformer = language / GenAI
• Parameter = learned by model; Hyperparameter = set by you
• Training = model learns; Inference = model predicts
• Real-time = sub-second; Batch = offline bulk; Async = large payload
• Ground Truth = labeling; Autopilot = AutoML; Clarify = bias + explainability
• Model Monitor = drift; Feature Store = reuse features
• JumpStart = pretrained in SageMaker; Pipelines = MLOps automation
• Bedrock = managed foundation models; Q = finished AI assistant
• Rekognition = images/video; Textract = documents/OCR
• Comprehend = NLP analysis; Kendra = enterprise search
• Transcribe = speech→text; Polly = text→speech; Translate = languages
• Lex = chatbots; Forecast = time-series; Personalize = recommendations
• Lookout for Metrics = KPI anomalies

🧠 Final Self-Test (10 questions)

Answer aloud before reading the answer:

1. A bank flags fraud where 1 in 1000 transactions is fraud. Best metric? Why is accuracy bad?

→ F1 / recall. Always-legit predicts 99.9% accurate but useless.

1. 50M customer records, segment without predefined groups. Type and problem?

→ Unsupervised, clustering.

1. 98% training, 71% validation. Diagnosis + 3 fixes?

→ Overfitting. More data, regularization, dropout, simpler model, early stopping.

1. Extract line items, totals, tax from PDF invoices. Service?

→ Amazon Textract.

1. Detect when input data shifted away from training data. Service?

→ SageMaker Model Monitor.

1. Score 50M customers Sunday night, no real-time need. Deployment?

→ Batch Transform.

1. Flight booking chatbot with departure / destination / date / passengers. Service?

→ Amazon Lex.

1. Difference between a parameter and a hyperparameter? One example each.

→ Parameter = learned by model (NN weights). Hyperparameter = set by you (learning rate).

1. Which deep learning architecture powers modern LLMs and why?

→ Transformers — attention lets every token see every other token, scales massively.

1. When pick Bedrock instead of SageMaker?

→ Bedrock = managed API for foundation models, no infra. SageMaker = custom training, full control.

If you got 8+ right, you're ready. Move to practice exams. If under 8, find the weak sections and reread.