AWS AIF-C01 — Day 1 Complete Study Guide

Before any topic, fix this mental model in your head:

Every exam question is testing one or more of these four layers. When you read a question, mentally walk down this ladder.

The "type" answers one question: What kind of data did the model learn from?

2.1 Supervised Learning

The intuition: Imagine teaching a child what a cat is by showing them 1,000 photos labeled "cat" and 1,000 photos labeled "not cat." The labels are the supervision. The child (model) learns to map photo → label.

Mechanics:

• Input data has a known correct answer (a label)
• Model learns the mapping input → output
• After training, give it a new input and it predicts the output

Two flavors of supervised learning:

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 in front of them with no labels and say: "Sort these into groups however you want." The child might separate by color, by animal type, by background. They find 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: Think of 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. There's no labeled dataset — there's a goal and feedback.

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. Semi-supervised learning uses both.

Mechanics:

• Small amount of labeled data + large amount of unlabeled data
• Model uses the unlabeled data to learn structure, and the labeled data to anchor the 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)

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

This is the lifecycle of every ML system. Memorize the order.

Stage 1 — Data Collection

Gather raw data: databases, logs, images, audio, IoT sensors, clickstream, public datasets.

Stage 2 — Preprocessing

Clean the raw data so the model can use it.

Stage 3 — Feature Engineering

Transform raw fields into useful features that help the model learn.

Why features matter: A model can't extract every pattern on its own. Smart features make patterns obvious.

Examples:

• Raw: date_of_birth = 1995-03-12 → Feature: age = 30
• Raw: transaction_timestamp → Features: hour_of_day, day_of_week, is_weekend
• Raw: address → Features: city, pincode, is_metro
• Text: convert to TF-IDF vectors, embeddings, etc.

Stage 4 — Training

The model learns by adjusting its internal parameters to minimize prediction error.

Stage 5 — Evaluation

Test the trained model on data it has never seen. This tells you if it actually generalizes or if it just memorized.

Metrics depend on the problem type — we'll cover these in detail in Topic 10.

Stage 6 — Deployment

Put the model into production so users / apps can use it.

Stage 7 — Monitoring

After deployment, watch for problems:

When monitoring detects problems, you retrain the model with new data. This is the loop.

5.1 Labeling

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

AWS service: SageMaker Ground Truth. It coordinates human labelers (mechanical turk-style) and uses ML to speed up the process (active learning + auto-labeling).

5.2 Data Splits

You split your dataset into 3 parts:

5.3 Class Imbalance

The problem: Fraud detection has 1% fraud, 99% legit. A naive classifier predicts "legit" always → 99% accuracy → useless.

Fixes:

5.4 Data Augmentation

Create modified versions of existing data to expand your training set artificially.

For images: rotate, crop, flip, change brightness, add noise, zoom.
For text: synonym replacement, back-translation, random word dropout.
For audio: time-stretching, pitch-shifting, adding background noise.

Purpose: Reduces overfitting and helps generalization. The model sees more "variety" without you collecting new data.

This distinction shows up constantly.

6.1 Parameters

Numbers the model learns automatically during training.

Examples:

• Weights in a neural network
• Biases
• Learned word embeddings
• Tree splits in a decision tree

You don't set these. The optimizer does.

6.2 Hyperparameters

Numbers YOU choose before (or during) training.

Examples:

• Learning rate
• Batch size
• Number of epochs
• Number of layers / neurons per layer
• Dropout rate
• Regularization strength
• Number of trees in a random forest
• Max tree depth

You tune these. This is called hyperparameter tuning — try different combinations and see which gives the best validation performance.

Memory hook

7.1 Training

The model learns.

7.2 Inference

The model predicts on new inputs.

7.3 Inference Modes (Critical for AWS exam)

This mapping is on the exam in many forms. Memorize it.

This is the #1 ML concept. The exam will test it many ways.

8.1 Overfitting

The intuition: A student who memorizes the answer key but doesn't understand the subject. Aces practice tests, fails the real exam.

The model: Learns the noise in the training data, not just the pattern. Performs great on training data, poorly on new data.

Signs:

• Training accuracy: 99%
• Test accuracy: 65%
• Big gap = overfitting

Causes:

• Model too complex (too many parameters for too little data)
• Not enough training data
• Trained for too many epochs
• Too many features (some are just noise)

Fixes (memorize these — exam favorite):

8.2 Underfitting

The intuition: A student who studied only the chapter titles. Bad at practice tests, bad at the real exam.

The model: Too simple to capture the pattern. Performs poorly on both training and test data.

Signs:

• Training accuracy: 60%
• Test accuracy: 58%
• Both are low

Causes:

• Model too simple
• Not enough features
• Training stopped too early
• Wrong algorithm choice

Fixes:

• Use a more complex model
• Add more / better features
• Train longer
• Reduce regularization (it might be too aggressive)

Closely related to overfitting/underfitting, but a deeper view.

9.1 Bias

Error from oversimplified assumptions. The model is too rigid to capture reality.

Example: Trying to fit a straight line to data that's shaped like a curve. No matter how much data you give it, a line can't bend.

High bias = underfitting.

9.2 Variance

Error from being too sensitive to the specific training set. Tiny changes in training data → wildly different model.

Example: A super-flexible squiggly curve that passes through every training point exactly. Slightly different data → totally different squiggle.

High variance = overfitting.

9.3 The Tradeoff

You can't easily have zero bias AND zero variance. Reducing one often increases the other.

The art of ML is finding the sweet spot.

This is where many candidates lose points. Master this section.

10.1 Setup — the Confusion Matrix

For binary classification, every prediction lands in one of 4 boxes:

Memorize this. All metrics derive from it.

10.2 Accuracy

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

Use when: Classes are roughly balanced (50/50 or close).

10.3 Precision

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:

• If you mark a legit email as spam (FP), the user misses something important. Bad.
• High precision = "When I say spam, it's really spam."

10.4 Recall (Sensitivity)

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

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

Classic example — Cancer screening:

• If you miss a real cancer case (FN), the patient doesn't get treated. Catastrophic.
• High recall = "I catch nearly every real cancer case, even if I have some false alarms."

10.5 The Precision–Recall Tradeoff

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

• Lower the threshold → flag more things as positive → catch more (↑ recall) but more false alarms (↓ precision)
• Raise the threshold → only flag the most certain cases → fewer false alarms (↑ precision) but miss some (↓ recall)

Choose based on what's costlier in your domain.

10.6 F1 Score

Plain English: Harmonic mean of precision and recall. A single number that balances both.

Use when:

• Imbalanced data
• Both false positives and false negatives matter
• You want one combined metric

10.7 AUC-ROC

ROC curve: Plots Recall vs False Positive Rate at every possible threshold.

AUC (Area Under Curve): A single number from 0 to 1.

• 1.0 = perfect classifier
• 0.5 = random guessing
• <0.5 = worse than random (and you should flip predictions)

Use when:

• Binary classification
• You want a threshold-independent measure
• Comparing multiple classification models

Plain English: "How well does the model separate positives from negatives, regardless of where I draw the line?"

10.8 Regression Metrics (briefly)

10.9 The Master Cheat Table — memorize this cold

You don't need to derive backprop. You need to know what each piece does.

11.1 The Neuron

A single neuron does:

• Inputs → numerical features
• Weights → learned numbers controlling each input's importance
• Bias → a learned offset (lets the neuron fire even when all inputs are 0)
• Activation function → adds non-linearity (more on this below)

11.2 Layers

11.3 Activation Functions

Without these, a neural network is just a glorified linear model. Activations let it learn curves and complex shapes.

11.4 Backpropagation (Backprop)

The intuition: The network makes a prediction. You compare it to the truth. The error gets propagated backward through the network, and each weight is nudged in the direction that would have reduced the error.

This is how a network learns.

Three families dominate the exam.

12.1 CNN — Convolutional Neural Network

Best for: images.

Why: A regular neural network treats each pixel independently. A CNN uses convolutional filters that detect local patterns (edges, textures, shapes) and stack them into higher-level concepts (eye → face).

Use cases:

• Image classification (cat vs dog)
• Object detection (find all cars in this photo)
• Facial recognition
• Medical imaging (tumor detection)
• Defect detection in manufacturing
• Self-driving perception

12.2 RNN — Recurrent Neural Network

Best for: sequential / time-ordered data.

Why: RNNs have a "memory" — output at time t depends on input at time t AND the hidden state from time t-1. This lets them process sequences.

Use cases:

• Time-series forecasting
• Speech recognition (audio over time)
• (Older) text generation
• Stock price prediction
• Sensor data over time

Weakness: Struggles with long-range dependencies. If a sentence is 50 words long, by the time RNN reaches word 50, it's largely forgotten word 1. Transformers fixed this.

12.3 Transformers

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

Why they dominate: Instead of processing tokens one at a time like RNNs, transformers use an attention mechanism that lets every token directly look at every other token. This means:

• They can handle very long sequences
• They train in parallel (massively faster than RNNs)
• They scale beautifully to billions of parameters

Use cases:

• Text generation
• Translation
• Summarization
• Question answering
• Chatbots
• Code generation
• All Large Language Models (LLMs)
• All Foundation Models

Quick mapping

MLOps = DevOps applied to ML. Building and shipping models reliably.

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 Monitoring (in production)

13.4 Retraining Triggers

You retrain when:

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

AWS service: SageMaker Pipelines automates this whole loop.

SageMaker AI is AWS's flagship ML platform. It covers the full ML lifecycle: prepare data, build, train, tune, deploy, monitor.

The exam will test which feature of SageMaker handles which part of the lifecycle. Memorize these.

14.1 SageMaker Studio

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

14.2 SageMaker Autopilot

• What it is: AutoML. Give it a dataset → it tries algorithms, tunes hyperparameters, and produces a deployable model.
• Use when: You want a model with minimal coding / minimal ML expertise.
• 🔑 Clue: "AutoML", "automatically build model", "least amount of code"

14.3 SageMaker Ground Truth

• What it is: Data labeling service (uses humans + ML to speed it up).
• 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: You need 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: You need fairness checks or explainability.
• 🔑 Clue: "bias detection", "explainability", "why did the model predict this?"

14.6 SageMaker Model Monitor

• What it is: Monitors deployed models for data quality issues, drift, and bias.
• Use when: You need ongoing monitoring of production models.
• 🔑 Clue: "detect drift", "monitor deployed model"

14.7 SageMaker JumpStart

• What it is: Library of pretrained models, foundation models, and solution templates.
• Use when: You want to start fast with a pretrained model instead of training from scratch.
• 🔑 Clue: "pretrained model", "quick start", "foundation model", "solution template"

14.8 SageMaker Feature Store

• What it is: A centralized repository for ML features (the engineered inputs you feed models).
• Use when: Multiple teams / models need to reuse the same features consistently for both training and inference.
• 🔑 Clue: "store and reuse features", "consistency between training and inference features"

14.9 SageMaker Data Wrangler (worth knowing)

• What it is: Visual data prep and feature engineering tool.
• Use when: Cleaning, transforming, and exploring data before training.

14.10 SageMaker Canvas (worth knowing)

• What it is: No-code ML for business analysts. Drag-and-drop model building.
• Use when: Non-technical users want to build ML models.

Three levels of control:

Common built-in algorithms:

Exam pattern:

These are API-based services. You don't train a model. You send data → AWS returns the AI result. Think of them as "AI as a service."

17.1 Amazon Rekognition — images & video

• Face detection, object detection, label detection
• Content moderation (detect unsafe content)
• Celebrity recognition
• Text in images (e.g., a sign in a photo)
• 🔑 Clue: image / video analysis, detect faces / objects

17.2 Amazon Comprehend — text understanding (NLP)

• Sentiment analysis (positive / negative / neutral)
• Entity detection (people, places, organizations)
• Key phrase extraction
• Language detection
• Topic modeling
• Document classification (custom)
• Comprehend Medical — medical entities (drug names, conditions)
• 🔑 Clue: understand text, sentiment, entities, key phrases

17.3 Amazon Textract — extract data from documents

• OCR + structure extraction
• Pulls out text, tables, forms, and key-value pairs
• Works on scanned PDFs, invoices, receipts
• 🔑 Clue: OCR, forms, tables, invoice extraction, document data extraction

17.4 Amazon Transcribe — speech → text

• Audio / video → text
• Real-time or batch
• Speaker identification, custom vocabulary
• Transcribe Medical for medical dictation
• 🔑 Clue: audio to text, speech recognition, call transcription, captions

17.5 Amazon Polly — text → speech

• Generates natural-sounding voice from text
• Many languages and voices, including neural voices
• 🔑 Clue: text to speech, natural voice, audio generation from text

17.6 Amazon Translate — language translation

• Real-time and batch translation between many languages
• 🔑 Clue: translate languages, multilingual

17.7 Amazon Lex — chatbots / voice bots

• Builds conversational interfaces using intents and slots
• Same engine as Alexa
• 🔑 Clue: chatbot, voice bot, conversational interface, IVR, intent and slots

18.1 Amazon Forecast — time-series forecasting

• Predicts future values from historical time-series
• Demand forecasting, inventory, workforce
• 🔑 Clue: forecast, time-series prediction, future demand

18.2 Amazon Personalize — recommendations

• Real-time personalized recommendations (same engine Amazon.com uses)
• "Users who bought X also bought Y," personalized ranking, similar items
• 🔑 Clue: recommend, personalize, user-specific suggestions

18.3 Amazon Kendra — intelligent enterprise search

• Natural-language search across company documents (S3, SharePoint, Confluence, etc.)
• Returns answers, not just links
• 🔑 Clue: enterprise search, search across documents, knowledge discovery, natural-language search

18.4 Amazon Fraud Detector — fraud detection

• Online payment fraud, fake account detection, risk scoring
• ⚠️ Check current AWS product status before relying on it in production. Existing customers may still use it, and the exam may still mention the concept. New fraud-detection use cases are often built on SageMaker instead.
• 🔑 Clue: fraud detection

18.5 Amazon Lookout for Metrics — anomaly detection in business metrics

• Detects unusual changes in KPIs (revenue, web traffic, conversion)
• 🔑 Clue: detect anomalies in metrics, business KPI monitoring, revenue / traffic spikes

Bonus services worth recognizing

This is the table you must know cold. Every exam has 5+ questions that map directly to this.

Trap 1 — "Accuracy is the right metric"

Almost never on the exam. If the dataset is imbalanced (fraud, disease, etc.) → use F1, Recall, or AUC-ROC.

Trap 2 — Rekognition vs Textract

• Rekognition = images/videos as visual content (faces, objects, scenes).
• Textract = images of documents where you want the text/tables/forms extracted.

Trap 3 — Transcribe vs Polly

• Transcribe = audio → text
• Polly = text → audio

Mnemonic: Transcribe writes down what was said. Polly reads things out loud.

Trap 4 — Comprehend vs Kendra

• Comprehend = analyze text (sentiment, entities, classification)
• Kendra = search across documents

Trap 5 — Forecast vs Personalize

• Forecast = predict numbers over time (sales next month)
• Personalize = recommend items to users

Trap 6 — When to use SageMaker vs a Pre-built AI service

Trap 7 — Real-time vs Batch vs Async Inference

• "low latency", "during checkout", "live" → Real-time
• "overnight", "millions of records", "no rush" → Batch
• "large input (>1MB or video)", "long processing", "async OK" → Async

Trap 8 — Underfitting fixes

A common wrong answer: "add regularization". That's the fix for overfitting. For underfitting → reduce regularization, add complexity, train longer.

Trap 9 — "Model is biased" usually means data is biased

The first answer is rarely "tune the model" — it's "collect more representative data" or "use SageMaker Clarify to detect and mitigate bias."

Trap 10 — Foundation Models

• JumpStart = pretrained models hosted in SageMaker (you can fine-tune)
• Bedrock = managed API for foundation models (Claude, Llama, Titan), no infrastructure

You will not fail this exam on definitions. You will fail it on service selection under pressure. So focus accordingly.

Recommended order for today:

1. Read Topics 1–9 in order (foundations: ML types → problem types → pipeline → training data → hyperparameters → inference vs training → over/underfitting → bias/variance). 30–45 min.
2. Drill Topic 10 (metrics) until you can answer "spam filter → precision," "cancer screening → recall," "imbalanced data → F1" without thinking. 20 min.
3. Read Topics 11–13 (neural networks, deep learning model types, MLOps). 30 min.
4. Memorize Topics 14–18 (SageMaker + AI services). This is the heaviest section. Use spaced repetition. 60+ min.
5. Memorize Topic 19 (the master Which-Service table) cold. Cover the right column with your hand and quiz yourself. Repeat until you get 100%. 30 min.
6. Read Topic 20 (traps) twice. Each trap is worth 1–2 exam points. 15 min.
7. Use Topic 21 as your final 5-minute pre-exam revision.

Self-test before you stop today

Answer these in your head. If you stumble, re-read the relevant topic.

1. A bank wants to flag suspicious transactions where 0.1% are fraud. Which evaluation metric should it prioritize?
2. A retailer has 50 million customer records and wants to find natural shopping segments without predefined labels. ML type? Problem type?
3. The model gets 99% training accuracy and 71% validation accuracy. Diagnosis and 3 fixes?
4. A company wants to extract line items and totals from PDF invoices. Which AWS service?
5. A team needs to monitor a deployed model for input distribution changes. Which SageMaker feature?
6. A model needs to score 50 million records every Sunday night, no real-time requirement. Which SageMaker deployment?
7. A startup wants a chatbot that handles "book a flight" with multiple slots (date, destination, passengers). Which AWS service?
8. Difference between a parameter and a hyperparameter? One example of each.
9. Which deep learning architecture powers modern LLMs and why?
10. When would you choose Amazon Bedrock vs SageMaker?

Answers:

1. Recall (or F1). Missing fraud is costly. Accuracy is misleading on imbalanced data.
2. Unsupervised, Clustering.
3. Overfitting. Fixes: more data, regularization, dropout, simpler model, early stopping (any 3).
4. Amazon Textract.
5. SageMaker Model Monitor (data drift).
6. Batch Transform.
7. Amazon Lex (intents + slots).
8. Parameters are learned by the model (e.g., neural network weights). Hyperparameters are set by you (e.g., learning rate).
9. Transformers — they use attention, scale to billions of parameters, and parallelize well.
10. Bedrock when you want a managed API for foundation models with no infra. SageMaker when you need to train a custom model or have full ML lifecycle control.

If you got 8+ correct, you're in great shape for Day 1. Don't try to do practice MCQs yet — let this material settle. Tomorrow, hit practice questions hard.