AWS Certified AI Practitioner — Task Statement 1.1 Study Notes

Explain Basic AI Concepts and Terminologies

This task statement tests whether you understand the fundamental vocabulary of artificial intelligence, machine learning, generative AI, data, model training, inferencing, and learning methods.

1. Artificial Intelligence Fundamentals

1.1 Artificial Intelligence — AI

Artificial intelligence (AI) is the field of computer science focused on creating systems that perform tasks normally associated with human intelligence.

Examples of human-like capabilities include:

Learning from data
Recognizing images and speech
Understanding language
Making predictions
Finding patterns
Generating content
Making decisions
Solving problems
Detecting anomalies

The overall goal of AI is to build systems that can derive useful meaning from data and perform tasks with limited explicit human instructions.

Examples of AI applications

Fraud detection
Medical image analysis
Product recommendations
Chatbots and virtual assistants
Demand forecasting
Resume screening
Predictive maintenance
Translation
Image and video analysis
Personalized content recommendations

Exam point

AI is the broadest term. It includes many approaches, including:

Rule-based systems
Machine learning
Deep learning
Generative AI
Reinforcement learning
Agentic AI

Not all AI systems learn from data. A simple rules-based expert system can be considered AI even though it does not use machine learning.

1.2 Machine Learning — ML

Machine learning (ML) is a subset of AI in which systems learn patterns from data and use those patterns to make predictions or decisions.

Instead of explicitly programming every rule, developers provide:

Data
An algorithm
Sometimes the expected answers or labels

The ML system adjusts its internal values during training so that it can make useful predictions on new data.

Example

An online store can use ML to recommend products based on:

Customer browsing history
Previous purchases
Product ratings
Similar customers’ behavior
Product characteristics

Key ML concepts

Term	Meaning
Feature	An input variable used by a model
Label/target	The expected output, mainly used in supervised learning
Algorithm	A mathematical or computational procedure used to learn patterns
Model	The trained result produced by applying an algorithm to data
Parameter	An internal value learned or adjusted during training
Training	The process of learning from data
Inference	Using a trained model to generate a prediction or output
Prediction	The output produced by a model
Model artifact	The saved trained parameters, model definition, and related metadata

Important distinction: algorithm versus model

An algorithm is the procedure or method used to learn.
A model is the trained result after the algorithm has processed data.

For example:

Linear regression is an algorithm.
A trained equation that predicts height from weight is a model.

1.3 Deep Learning

Deep learning is a specialized type of machine learning that uses multilayer neural networks.

Deep learning models are particularly useful for complex and high-dimensional data, such as:

Images
Video
Audio
Speech
Text
Natural language

Neural networks

A neural network is made up of connected computational units called nodes, inspired loosely by biological neurons.

A typical neural network contains:

Input layer
Receives the input features.
Hidden layers
Transform the input and learn increasingly complex patterns.
Output layer
Produces the prediction or result.

The model assigns weights to inputs. During training, these weights are repeatedly adjusted to reduce the difference between the predicted output and the expected output.

Why deep learning is useful

Traditional ML often requires people to manually select and extract useful features. Deep learning can learn many features automatically.

For example, in image recognition:

Early layers may learn edges.
Middle layers may learn shapes.
Later layers may identify objects such as fish, cars, or faces.

Traditional ML versus deep learning

Characteristic	Traditional ML	Deep learning
Typical data	Structured/tabular data	Images, video, audio, text
Feature engineering	Often requires manual feature selection	Learns features automatically
Data requirement	Usually lower	Often requires very large datasets
Compute requirement	Generally lower	Generally higher
Interpretability	Often easier	Often more difficult
Common use cases	Churn, recommendations, classification	Computer vision, NLP, speech

Exam trap

Deep learning is not separate from ML. It is a subset of ML, and ML is a subset of AI.

The relationship is:

AI ⟶ Machine Learning ⟶ Deep Learning

2. Computer Vision and Natural Language Processing

2.1 Computer Vision

Computer vision is the use of AI and ML to process and understand images and video.

Common computer vision tasks include:

Image classification
Object detection
Facial recognition
Image segmentation
Optical character recognition
Defect detection
Image moderation
Activity recognition
Anomaly detection

Common computer vision outputs

A model might return:

The class of an image
A probability or confidence score
The coordinates of an object
A bounding box
A segmentation mask
Detected text
Whether an image contains unsafe content

Example

A manufacturing system can inspect circuit boards and detect:

Scratches
Missing components
Incorrect assembly
Surface defects

2.2 Natural Language Processing — NLP

Natural language processing (NLP) enables computers to understand, interpret, translate, analyze, and generate human language.

NLP applications include:

Chatbots
Virtual assistants
Language translation
Text classification
Sentiment analysis
Text summarization
Question answering
Speech-to-text
Text-to-speech
Entity extraction
Document analysis

For example, an NLP system can understand that a customer’s message expresses dissatisfaction even if the customer does not use the exact word “complaint.”

NLP and generative AI

NLP is a broad field. Generative AI is one type of technology that can be used for NLP tasks, such as:

Generating an answer
Summarizing a document
Translating text
Writing an email

3. Algorithms, Models, Training, and Inference

3.1 Algorithm

An algorithm is a defined mathematical or computational method for solving a problem or learning a relationship in data.

Examples include:

Linear regression
Logistic regression
Decision trees
Random forests
K-means clustering
Neural networks
Transformer architectures

An algorithm determines how data is processed and how a model learns.

3.2 Features

A feature is an input property used by a model.

Examples:

Customer age
Income
Number of previous purchases
Temperature
CPU utilization
Pixels in an image
Words or tokens in a document

In a tabular dataset, features are often represented as columns.

Example

Weight	Age	Height
70 kg	30	175 cm

Weight and age could be features.
Height could be the label or target.

3.3 Labels

A label is the known correct output associated with a training example.

Examples:

“Fraud” or “Not fraud”
“Fish” or “Not fish”
House price
Customer churn: “Yes” or “No”
Disease diagnosis
Customer sentiment

Labels are required for traditional supervised learning.

3.4 Training

Training is the process of using data to adjust a model’s internal parameters so that it produces useful outputs.

A simplified training process is:

Provide training data to an algorithm.
The model produces a prediction.
Compare the prediction with the known answer, if available.
Calculate the error or loss.
Adjust model parameters.
Repeat until the model performs sufficiently well.

Model parameters

Parameters are values learned during training.

For linear regression:

\[ y = mx + b \]

\(x\) is the input.
\(y\) is the output.
\(m\) is the slope.
\(b\) is the intercept.

The values of \(m\) and \(b\) are adjusted during training to find the best-fitting line.

3.5 Model Artifacts

Training typically produces model artifacts, which may contain:

Learned parameters
Model architecture or definition
Preprocessing information
Configuration
Metadata

These artifacts are commonly stored in Amazon S3 and packaged with inference code to create a deployable model.

AWS exam connection

Amazon S3 is commonly used as a source or destination for ML data because it offers:

High durability
Large-scale storage
Support for many data formats
Integration with AWS ML services
Cost-effective object storage

3.6 Inference

Inference is the process of using a trained model to produce an output from new data.

Examples:

Predict whether a transaction is fraudulent.
Classify an image as containing a fish.
Forecast next month’s sales.
Generate a response to a prompt.
Estimate the probability that a customer will leave.

Important point

An inference is generally a prediction or probabilistic result, not guaranteed truth.

A classification model might return:

Fish: 0.93 probability
Not fish: 0.07 probability

The model output is based on learned patterns and may be incorrect.

4. Types of Inferencing

The exam may ask you to select an inference method based on latency, request volume, cost, and whether results are needed immediately.

4.1 Real-Time Inference

Real-time inference uses a persistent endpoint that is continuously available to receive requests.

Characteristics

Low latency
Immediate response
Persistent compute resources
Suitable for interactive applications
Handles a sustained flow of requests

Use cases

Fraud detection during a transaction
Online recommendations
Chatbot responses
Image classification in an application
Predicting customer churn while a user is online

AWS examples

Amazon SageMaker real-time endpoints
Amazon Bedrock model invocation
Applications using an API that requires an immediate response

Exam clue

If the question says:

“Low latency”
“Immediate response”
“Online prediction”
“Interactive application”
“Endpoint always available”

The answer is usually real-time inference.

4.2 Batch Inference

Batch inference processes a large volume of data at once, usually on a schedule.

Characteristics

Data is available upfront.
Results do not need to be returned immediately.
Compute runs for the batch job and then stops.
Often more cost-effective for large offline workloads.
No continuously available endpoint is required.

Use cases

Monthly sales forecasting
Scoring an entire customer database
Generating recommendations overnight
Processing historical data
Producing a daily report
Predicting inventory requirements

AWS example

Amazon SageMaker Batch Transform can generate predictions for an entire dataset without deploying a persistent endpoint.

Exam clue

If the question says:

“Offline”
“Large dataset”
“Run once per day or month”
“Results can wait”
“Cost optimization”
“No persistent endpoint”

The answer is usually batch inference.

4.3 Asynchronous Inference

Asynchronous inference is appropriate when requests may take a long time to process or contain large input payloads, and the client does not need an immediate response.

The client submits a request and receives the result later.

Typical workflow

Client submits input.
The request is placed in a queue.
The model processes the request.
The output is stored, commonly in Amazon S3.
The client is notified or checks for the result.

Use cases

Large image or video processing
Long-running document analysis
Large text generation requests
Requests with unpredictable processing time
Applications that can tolerate delays

Exam clue

If the question describes:

A request that takes time
A response that can be returned later
Large payloads
Queue-based processing
Notifications when processing finishes

The answer is likely asynchronous inference.

4.4 Serverless Inference

Serverless inference runs inference without requiring the customer to manage or maintain dedicated infrastructure.

Resources are provisioned when needed and may scale down when idle.

Characteristics

No server management
Suitable for intermittent or unpredictable traffic
Can reduce cost when the endpoint is not constantly used
May have startup or cold-start latency
Not always ideal for strict, consistently low-latency requirements

AWS examples

Amazon SageMaker Serverless Inference
Amazon Bedrock managed model invocation

Exam trap

“Serverless” describes how infrastructure is managed and billed. It is not necessarily the same as batch or real-time.

For example:

Serverless inference can support an on-demand request.
Batch inference can use temporary compute.
Real-time inference can be serverless in some architectures, although persistent low latency may favor provisioned resources.

4.5 Comparison of Inference Types

Type	Response timing	Infrastructure	Best for
Real-time	Immediate	Persistent endpoint	Interactive, low-latency requests
Batch	Later, after job completion	Runs during batch	Large offline datasets
Asynchronous	Later per request	Queue-based or managed processing	Long-running or large requests
Serverless	Depends on service and invocation	Managed, automatically provisioned	Intermittent or unpredictable traffic

5. Data Types Used in AI and ML

5.1 Structured Data

Structured data follows a defined schema and is usually organized in rows and columns.

Examples:

CSV files
Relational database tables
Sales records
Customer records
Financial transactions
Inventory data

AWS examples

Amazon RDS
Amazon Redshift
Amazon S3 containing CSV files

Structured data is commonly queried using SQL.

5.2 Semi-Structured Data

Semi-structured data does not follow a strict table structure but contains tags, keys, or metadata that provide organization.

Examples:

JSON
XML
Log files
Event records
Web data

Attributes may be missing or vary between records.

AWS examples

Amazon DynamoDB
Amazon DocumentDB
Amazon S3 storing JSON documents

5.3 Unstructured Data

Unstructured data does not follow a predefined data model or tabular schema.

Examples:

Images
Video
Audio
Documents
Emails
Social media posts
Free-form text

Unstructured data is commonly stored as objects in Amazon S3.

ML features may need to be extracted from unstructured data using techniques such as:

Tokenization for text
Image preprocessing
Speech recognition
Embeddings
Object detection
Feature extraction

5.4 Labeled and Unlabeled Data

Labeled data

Labeled data includes both:

Input features
Expected output

Example:

Image	Label
Image of fish	Fish
Image of dog	Not fish

Labeled data is used primarily in supervised learning.

Unlabeled data

Unlabeled data contains input data but no known target output.

Examples:

Customer transactions without fraud labels
Images without object labels
Network traffic without incident labels

Unlabeled data is used in unsupervised learning and can also be used in some self-supervised and generative AI approaches.

5.5 Tabular Data

Tabular data is organized into rows and columns.

Examples:

Customer churn data
Loan applications
Product catalogs
Sales history
Employee records

Tabular data is often structured, but tabular data and structured data are not exactly identical concepts. Tabular data describes the representation, while structured data describes how consistently the data follows a schema.

5.6 Time-Series Data

Time-series data consists of observations recorded in sequence and associated with timestamps.

Examples:

Stock prices
CPU utilization
Temperature readings
Sales per day
Website traffic
Transactions per second
Sensor readings

Time-series data is useful for:

Forecasting
Capacity planning
Demand prediction
Predictive maintenance
Anomaly detection

Example

Historical CPU usage can be used to predict when additional infrastructure capacity will be required.

Exam clue

If the question mentions:

Historical values
Dates or timestamps
Forecasting future values
Seasonal patterns
Sequential measurements

The answer likely involves time-series data or regression/forecasting.

5.7 Image, Video, Audio, and Text Data

These are generally considered unstructured data.

Images and video

Used for:

Defect detection
Object identification
Facial recognition
Content moderation
Medical analysis

Text

Used for:

Sentiment analysis
Classification
Summarization
Search
Translation
Question answering

Audio

Used for:

Transcription
Speaker identification
Voice assistants
Call-center analysis

6. Machine Learning Learning Methods

6.1 Supervised Learning

In supervised learning, the model trains on labeled data.

The model learns the relationship between:

Input features
Known target outputs

Common supervised learning tasks

Classification

Predicts a category or class.

Examples:

Fraud or not fraud
Fish or not fish
Spam or not spam
Approved or rejected
Positive, negative, or neutral sentiment

Regression

Predicts a numerical value.

Examples:

Sales amount
House price
Temperature
Number of customers
Future demand

Supervised learning workflow

Collect labeled examples.
Provide features and expected outputs.
Train the model.
Measure errors.
Adjust parameters.
Use the trained model for predictions.

Main challenge

Obtaining high-quality labels can be expensive and time-consuming.

AWS connection

Amazon SageMaker Ground Truth helps create labeled datasets. It can use human workers and can integrate with Amazon Mechanical Turk for crowdsourced labeling.

Exam trap

Supervised learning is not defined by whether a human is directly involved during prediction. It is defined by the use of labeled training data.

6.2 Unsupervised Learning

In unsupervised learning, the model receives data without predefined labels.

The model attempts to discover hidden structure or patterns.

Common unsupervised tasks

Clustering
Grouping similar data
Anomaly detection
Dimensionality reduction
Pattern discovery

Examples

Grouping customers by purchasing behavior
Identifying unusual network traffic
Detecting an abnormal sensor reading
Discovering different types of users
Organizing documents into topics

Clustering

Clustering divides data into groups based on similarity.

For example, a retailer might group customers into segments such as:

Occasional buyers
Frequent buyers
High-value customers
Discount-focused customers

Anomaly detection

Anomaly detection identifies observations that differ significantly from normal patterns.

Examples:

Unusual credit-card activity
A sensor reporting an abnormal temperature
A sudden drop in application traffic
A suspicious network event

Exam trap

Unsupervised learning has no specified correct output during training. It does not mean the model has no objective at all; it means the objective is usually to discover structure rather than match labeled answers.

6.3 Reinforcement Learning

Reinforcement learning (RL) trains an agent to make decisions by interacting with an environment.

The agent:

Observes the current state.
Takes an action.
Receives a reward or penalty.
Learns which actions are more likely to achieve its goal.

Key RL terms

Term	Meaning
Agent	The decision-making system
Environment	The world or system in which the agent acts
Action	A choice made by the agent
State	The current condition of the environment
Reward	Feedback indicating how desirable an action was
Policy	A strategy for selecting actions

Example: AWS DeepRacer

Agent: The race car
Environment: The racetrack
Action: Steering, accelerating, or slowing down
Goal: Stay on the track and complete the course
Reward: Positive feedback for staying on track and moving efficiently

Exploration versus exploitation

An RL agent must balance:

Exploration: Trying new actions to discover whether they are useful.
Exploitation: Choosing actions already known to produce good rewards.

Difference from unsupervised learning

Both can operate without labeled examples, but they are different:

Unsupervised learning	Reinforcement learning
Finds patterns in data	Learns through interaction
No predefined target label	Has a goal or reward objective
Often clusters or detects anomalies	Chooses actions
No reward feedback required	Uses rewards and penalties

Exam trap

Reinforcement learning is not simply “learning without labels.” The defining characteristics are:

An agent
An environment
Actions
Rewards or penalties
A goal

7. Model Fit: Overfitting and Underfitting

7.1 Model Fit

Fit describes how well a model represents the relationship between its input data and expected output.

A good model should generalize: it should perform well not only on training data but also on new, unseen data.

7.2 Overfitting

Overfitting occurs when a model performs very well on training data but poorly on new data.

The model has learned the training examples too specifically, including noise or irrelevant details.

Example

A fish classifier is trained mostly on fish swimming underwater. It performs well on those images but fails to recognize a fish outside water.

Causes

Too little training data
Training data is not diverse
Model is too complex
Training for too long
Model learns noise
Data leakage in the training process

Possible solutions

Use more diverse training data.
Add more training examples.
Use regularization.
Simplify the model.
Stop training earlier.
Use data augmentation.
Evaluate using unseen validation and test data.

Exam clue

High training accuracy + low test accuracy = overfitting.

7.3 Underfitting

Underfitting occurs when a model is too simple or insufficiently trained to learn the underlying relationship.

It performs poorly on both:

Training data
New data

Causes

Model is too simple
Insufficient training
Too few features
Dataset is too small
Excessive regularization
Poor feature selection

Possible solutions

Train for longer.
Use a more complex model.
Add useful features.
Obtain more representative data.
Reduce excessive regularization.

Exam clue

Low training accuracy + low test accuracy = underfitting.

7.4 Generalization

Generalization is the ability of a model to perform well on data it has not seen before.

The objective is not to memorize the training data. The objective is to learn patterns that apply to new examples.

8. Bias and Fairness

8.1 Bias in AI Models

Bias occurs when a model produces systematically skewed or unequal results, especially across different groups.

Bias may cause a model to favor or disadvantage a particular class or population.

Example

A loan approval model trained on data that lacks approved applications from a particular demographic group may incorrectly learn that the group should not receive loans.

Sources of bias

Bias can enter at many stages:

Biased data collection
Underrepresentation of groups
Historical discrimination in training data
Incorrect labels
Poor feature selection
Proxy variables
Sampling errors
Biased evaluation metrics
Human assumptions in system design

Proxy variables

A feature may act as a proxy for a protected characteristic even if the protected feature is removed.

For example:

Postal code may correlate with race or socioeconomic status.
School attended may correlate with demographic background.

Simply removing a field such as gender does not automatically eliminate bias.

8.2 Fairness

Fairness means that a model’s results should not unjustifiably disadvantage individuals or groups based on protected or sensitive characteristics.

Fairness must be considered throughout the ML lifecycle:

Define fairness requirements.
Inspect training data.
Identify representation gaps.
Evaluate model performance across groups.
Monitor deployed predictions.
Adjust the data, model, or decision process when necessary.

Important point

Fairness is not automatically guaranteed by removing sensitive attributes. Other inputs may still contain indirect information about those attributes.

AWS exam connection

AWS services and practices can support responsible AI through:

Dataset analysis
Bias detection
Model evaluation
Human review
Monitoring
Documentation
Governance controls

For this task statement, focus primarily on understanding the concepts of bias and fairness rather than memorizing a specific service.

9. Generative AI

9.1 Definition

Generative AI (GenAI) is AI that creates new content based on patterns learned from training data.

It can generate:

Text
Images
Audio
Video
Music
Code
Summaries
Conversational responses

Example

A user provides a prompt such as:

“Write a product description for a waterproof hiking backpack.”

The model generates an original response based on its learned patterns.

9.2 Prompts and Responses

A prompt is the input provided to a generative AI model.

The output may be called:

Response
Completion
Generated content
Model output

Prompt quality affects the usefulness of the result. Clear instructions, context, constraints, and examples often improve results.

9.3 Large Language Models — LLMs

A large language model (LLM) is a deep learning model trained on very large collections of text data.

LLMs learn statistical relationships between tokens and can perform tasks such as:

Text generation
Summarization
Translation
Question answering
Classification
Code generation
Information extraction
Conversational interaction

Tokens

A token is a unit of text processed by a language model. A token may be:

A whole word
Part of a word
Punctuation
A symbol

LLMs generate outputs by predicting likely next tokens based on the input and prior generated tokens.

Transformers

Modern LLMs commonly use transformer neural networks.

Transformers are effective because they can examine relationships between elements of a sequence using mechanisms such as attention.

Important clarification

Transformers can process many input tokens in parallel during training, which is more efficient than processing every input word strictly one at a time. During generation, however, output tokens are generally produced sequentially.

Model parameters

LLMs may contain billions of learned parameters. These parameters capture statistical patterns from training data. They do not represent a literal database of facts or guaranteed human understanding.

9.4 Generative AI versus Traditional Predictive ML

Characteristic	Traditional predictive ML	Generative AI
Main output	Prediction, classification, score	New content
Example output	Fraud probability: 0.97	Written fraud explanation
Common training	Often task-specific	Often pre-trained on broad datasets
Input	Features	Prompt or multimodal input
Typical use	Forecasting, classification	Text, image, code, audio generation
Output certainty	Often probability or numeric value	Generated and potentially variable

Exam trap

Generative AI does not mean the output is necessarily correct, original in a legal sense, or free of bias. Generated content still requires evaluation and appropriate safeguards.

10. Agentic AI

10.1 Definition

Agentic AI refers to AI systems that can pursue goals by planning, reasoning, selecting actions, using tools, and adapting based on results.

An agentic system may:

Receive a goal.
Break the goal into steps.
Decide what action to take.
Use external tools or data sources.
Observe the result.
Continue, revise, or stop.

Example

A travel-planning agent could:

Understand a user’s travel requirements.
Search flight and hotel systems.
Compare options.
Ask for clarification.
Create an itinerary.
Book a reservation after receiving approval.

Components commonly associated with agentic AI

Foundation model or LLM
Planning and reasoning
Memory or state
Tool/API access
Workflow orchestration
Guardrails
Human approval
Feedback and monitoring

Agentic AI versus generative AI

Generative AI	Agentic AI
Generates content or responses	Pursues a goal and takes actions
Often responds to a single prompt	May perform multiple steps
Usually produces an answer	May use tools and interact with systems
Does not necessarily act independently	Has some degree of autonomy
Example: write an email	Example: send the email after checking the calendar

Exam trap

An LLM chatbot is not automatically an agent. It becomes more agentic when it can:

Plan multiple steps
Invoke tools
Access external systems
Take actions
Maintain state
Work toward a goal with limited human intervention

Agentic AI is still AI and may use ML, deep learning, LLMs, and GenAI. These terms describe different aspects of the system.

11. Relationships and Differences Between Key AI Terms

11.1 Hierarchy

A useful hierarchy is:

Artificial Intelligence
└── Machine Learning
    └── Deep Learning
        └── Some Generative AI models, including many LLMs

Agentic AI is better viewed as a system behavior or application architecture. It may use generative AI and LLMs, but it is not simply another layer in the same hierarchy.

11.2 Summary comparison

Term	What it describes	Typical capability
AI	Broad field of intelligent computer systems	Reasoning, perception, decision-making
ML	AI that learns patterns from data	Predictions and classifications
Deep learning	ML using multilayer neural networks	Complex image, speech, and text tasks
GenAI	AI that creates new content	Text, images, code, audio, video
LLM	Large deep learning model trained primarily on text	Language understanding and generation
Agentic AI	Goal-oriented AI system that plans and acts	Tool use, multi-step execution, autonomous actions

Critical exam points

AI is the broadest category.
ML is a subset of AI.
Deep learning is a subset of ML.
GenAI commonly uses deep learning but is defined by its ability to generate content.
LLMs are a type of model, generally based on deep learning and transformers.
Agentic AI is defined by goal-directed behavior and action, not merely by generating text.

12. Exam Tips and Common Traps

Tip 1: Look for the defining characteristic

Do not select an answer based only on a familiar keyword.

Labeled data → supervised learning
No labels and grouping → unsupervised learning
Agent, environment, reward → reinforcement learning
New content generation → generative AI
Goal, planning, tools, actions → agentic AI
Images/video → computer vision
Human language → NLP

Tip 2: Distinguish training from inference

Training changes or learns model parameters.
Inference uses the trained model to produce an output.

A model is not retrained every time it makes a prediction.

Tip 3: Distinguish batch, real-time, and asynchronous inference

Immediate response → real-time
Large scheduled dataset → batch
Response later for a long-running request → asynchronous
No infrastructure management → serverless

These categories may overlap. For example, an application may use a serverless architecture for asynchronous processing.

Tip 4: Remember that predictions are probabilistic

An inference is usually not a guaranteed fact. Classification models frequently return probabilities or confidence scores.

A model can be highly accurate and still produce incorrect predictions.

Tip 5: Identify overfitting from training and test behavior

Good training performance, poor new-data performance → overfitting
Poor training and poor new-data performance → underfitting
Good performance on both → better generalization

Tip 6: Do not confuse bias with variance or randomness

For this exam context, bias generally refers to systematic disparities or unfair outcomes across groups.

A model can be accurate overall but still unfair to a particular group.

Tip 7: Structured and labeled are different concepts

These terms describe different properties:

Structured/unstructured describes data format.
Labeled/unlabeled describes whether expected outputs are provided.

Examples:

A CSV file can be structured and labeled.
An image dataset can be unstructured and labeled.
Customer records can be structured and unlabeled.
Text documents can be unstructured and unlabeled.

Tip 8: Deep learning is not automatically better

Deep learning may provide better results for complex unstructured data, but it often requires:

More data
More compute
More training time
More operational complexity
More difficult interpretation

Traditional ML may be more efficient for structured tabular data.

Tip 9: Generative AI does not guarantee factual accuracy

Generative AI can produce fluent and convincing content that is incorrect. This is commonly associated with hallucination.

The model generates likely content; it does not inherently verify truth.

Tip 10: Agentic AI requires more than conversation

A system that only answers questions is generally generative AI.

An agentic system can typically:

Plan
Use tools
Make decisions
Take actions
Coordinate multiple steps
Operate toward a goal

13. Rapid Review Cheat Sheet

AI: Broad field of creating intelligent systems.
ML: AI that learns patterns from data.
Deep learning: ML based on multilayer neural networks.
Computer vision: AI for images and video.
NLP: AI for understanding and generating human language.
Algorithm: Method used to learn or solve a problem.
Model: Trained result of applying an algorithm to data.
Training: Learning parameters from data.
Inference: Using a trained model to make a prediction or generate output.
Feature: Model input.
Label: Known expected output.
Overfitting: Performs well on training data, poorly on new data.
Underfitting: Performs poorly on both training and new data.
Bias: Systematic unfair or skewed outcomes.
Fairness: Avoiding unjustified disparities across groups.
GenAI: Produces new content.
LLM: Large deep learning model for language.
Agentic AI: Goal-oriented AI that plans, uses tools, and takes actions.
Real-time inference: Immediate response.
Batch inference: Large offline workload.
Asynchronous inference: Result returned later.
Serverless inference: Managed infrastructure without server administration.
Supervised learning: Uses labeled data.
Unsupervised learning: Finds patterns in unlabeled data.
Reinforcement learning: Agent learns actions through rewards and penalties.