Easy ReviewAIF-C01
Fundamentals of AI and ML27

AI and ML fundamentals

What AI, machine learning, deep learning, generative AI and agentic AI mean, and which kind of learning fits which problem.

The words the AI Practitioner exam builds on: how AI, machine learning, deep learning and generative AI nest inside each other, the three ways a model learns, and the problem types each one solves.

AI, ML, deep learning and generative AI

Artificial intelligence (AI) is the broad goal of making machines solve problems the way people do. Machine learning (ML) is the part of AI where a system learns from data; deep learning is ML with many-layered neural networks; generative AI uses deep learning to create new content. Picture four nested circles.

How it works

  1. AI covers any technique that makes a machine act intelligently, including hand-written rules.
  2. Machine learning replaces rules with a model trained on data: the algorithm finds the patterns, and the trained model makes predictions on new data.
  3. Deep learning uses neural networks with many layers of nodes, which can learn from images, audio and text directly.
  4. Generative AI uses large deep learning models to produce new text, images, audio or video instead of only labeling data.
  5. Agentic AI adds autonomy: the system plans and takes actions toward a goal, often calling tools along the way.
Branch Works on Example
Natural language processing (NLP) Text and language Sentiment of reviews
Computer vision Images and video Content moderation
Speech recognition Audio Call transcripts

Exam cues

  • If the question asks for the hierarchy, think AI > ML > deep learning > generative AI, because each is a subset of the one before.
  • If the question says the model creates new content, think generative AI; if it assigns labels, think a discriminative (traditional) model.
  • If the question says the system decides its next steps and uses tools on its own, think agentic AI.

Remember

  • Neural networks are layers of connected nodes; “deep” means many layers.
  • Foundation models and large language models are deep generative models.

Trap: “AI” and “machine learning” aren’t synonyms. A system that follows fixed rules is AI but involves no learning.

From the AWS docs (8)
  • Artificial intelligence is an umbrella term for techniques that let machines solve problems in human-like ways; machine learning and deep learning fall under it, but not all AI is machine learning.
  • Machine learning is one branch of AI that develops algorithms and statistical models which learn patterns from data instead of following explicit rules.
  • Deep learning is machine learning with neural networks of many layers, which learn complex patterns in images, text and sound.
  • An artificial neural network is made of artificial neurons (nodes) that process information together, loosely mirroring how the brain works.
  • Generative AI is the subset of AI that creates new content such as text, images, video and audio; deep generative learning is the basis of modern generative AI and foundation models.
  • Natural language processing interprets and generates human language, computer vision extracts information from images and video, and speech recognition turns speech into text; all three rely on deep learning.
  • Agentic AI is an autonomous AI system that acts independently to reach a goal, planning and carrying out complex tasks without step-by-step prompting.
  • Traditional discriminative models classify data, predicting a label from features; generative models learn the distribution of the data so they can create new samples, such as animal images that were not in the training set.

Supervised, unsupervised and reinforcement learning

A model learns in one of three ways: from labeled examples (supervised), by finding structure in unlabeled data (unsupervised), or from rewards for its actions (reinforcement). The data you have and the answer you want decide which one fits.

How it works

  1. Supervised: every training example has the right answer. A categorical answer is classification; a number is regression.
  2. Unsupervised: no answers in the data. The model groups (clustering), compresses (dimension reduction) or spots outliers (anomaly detection).
  3. Reinforcement: an agent acts in an environment, gets rewards or penalties, and learns the actions that pay off over time.
  4. Self-supervised: foundation models create their own training signal from unlabeled text or images, such as predicting the next word.
Problem Target Learning type
Binary classification One of two classes Supervised
Multiclass classification One of several classes Supervised
Regression A number Supervised
Clustering No target, find groups Unsupervised
Anomaly detection No target, find outliers Unsupervised

Exam cues

  • If the question says labeled historical data and a yes/no or category answer, think classification, because the target is categorical.
  • If the question says predict an amount, price or count, think regression, because the target is continuous.
  • If the question says group customers with no predefined groups, think clustering, because there are no labels.
  • If the question says rewards, penalties or trial and error, think reinforcement learning.

Remember

  • Labels present: supervised. Labels absent: unsupervised.
  • Clustering, dimension reduction and anomaly detection are unsupervised.

Trap: Predicting one of several categories is still classification (multiclass), not regression, even when the categories are numbered.

From the AWS docs (9)
  • Supervised learning trains on data whose inputs come with target values (labels).
  • When the target is categorical the problem is classification, either binary (one of two classes) or multiclass (one of several classes).
  • When the target is a continuous number the problem is regression, such as predicting a house price from its features.
  • Unsupervised learning works on data without labels and discovers patterns such as groupings in it.
  • Unsupervised problem types include clustering (grouping similar items), dimension reduction (keeping the most relevant features), anomaly detection (finding rare items such as bank fraud) and density estimation.
  • In reinforcement learning an agent learns by trial and error in an environment, choosing actions to maximize its long-term reward.
  • Foundation models are trained on broad, unlabeled data, learning to predict the next item in a sequence instead of learning from human labels.
  • Semi-supervised learning combines a small amount of labeled data with a large amount of unlabeled data: a model partly trained on the labeled data labels the rest (pseudo-labeling), and training continues on both.
  • The SageMaker AI k-nearest neighbors (k-NN) algorithm classifies a sample with the most frequent label among the k closest points, or for regression returns the average of their values.

Types of data for AI models

The kind of data you have shapes which model you can build. Supervised learning needs labeled data; text, images and audio are unstructured; sales by day are time series.

How it works

Data type Example Typical models or services
Structured (tabular) Customer table Classification, regression
Time series Hourly sensor readings Forecasting
Text Reviews, emails NLP, Amazon Comprehend, LLMs
Image and video Product photos Computer vision, Amazon Rekognition
Audio Call recordings Amazon Transcribe
  1. Labeled data has the answer attached and feeds supervised learning; unlabeled data feeds unsupervised and self-supervised learning.
  2. Labels usually come from people judging raw data, so labeling quality sets the ceiling for model quality.

Exam cues

  • If the question says rows and columns with a fixed schema, think structured (tabular) data.
  • If the question says readings over time or forecasting, think time series data.
  • If the question says the model learns badly from inconsistent labels, think label quality and ground truth.

Remember

  • Structured = predefined schema; unstructured = no schema.
  • Ground truth = the correctly labeled reference dataset.

Trap: A JSON log or a CSV doesn’t make the content structured for ML on its own; a column of free-text comments is still unstructured text.

From the AWS docs (5)
  • Structured data follows a predefined schema of field names and types and is usually tabular, in rows and columns.
  • Unstructured data, such as free text, images, audio and video, has no predefined data model and needs techniques such as natural language processing to analyze.
  • Data labeling adds meaningful labels to raw data, such as whether a photo contains a bird, so a supervised model can learn from it; a correctly labeled reference dataset is called ground truth.
  • A trained model can only be as accurate as its ground truth, so accurate labeling is essential.
  • Time series data is a sequence of data points recorded over a time interval, such as a stock price, an instance's CPU use or an IoT sensor's temperature over time.

Common algorithms and model families

Classic ML algorithms each fit a problem type, and neural network families each fit a data type. Knowing the pairing answers many “which technique” questions.

How it works

Algorithm or family Problem Example
Linear regression Regression House price
Logistic regression Classification with probabilities Churn yes or no
Decision trees, XGBoost Classification or regression on tables Fraud score
K-means Clustering Customer segments
CNN Images Image classification, object detection
RNN Sequences Time-series forecasting
Transformer Language and more LLMs

Exam cues

  • If the question says images, think CNN; sequences over time, think RNN; language at scale, think transformer.
  • If the question says segments with no labels, think k-means clustering.
  • If the question says yes/no with a probability, think logistic regression.
  • If the question says locate objects in an image, think object detection.

Remember

  • Logistic regression classifies; linear regression predicts numbers.
  • Multi-label: several labels per item; multi-class: one of several.

Trap: Decision trees are easy to explain but a deep ensemble of trees (XGBoost) is much harder to interpret.

From the AWS docs (8)
  • The SageMaker AI Linear Learner algorithm learns a linear function for regression or a linear threshold function for classification.
  • Logistic regression is a classification algorithm that returns probabilities, which are compared against a threshold to assign a class.
  • XGBoost is a gradient-boosted trees algorithm that combines an ensemble of simpler, weaker models.
  • Simpler models such as decision trees and linear regression are more transparent and easier to interpret than deep neural networks.
  • K-means finds discrete groupings in data, where members of a group are as similar as possible to each other and as different as possible from other groups, such as high-, medium- and low-spending customers.
  • Convolutional neural networks (CNNs) use hidden layers that apply convolutions, which makes them very useful for image classification and recognition.
  • Recurrent neural networks (RNNs) process inputs one step at a time in sequence and are used for sequences such as time-series forecasting, as in the SageMaker AI DeepAR algorithm.
  • Multi-label image classification tags an image with every label that applies, and object detection finds and classifies people and objects within an image.

Flashcards

tap to flip
Answer

·

All 27 cards as a list

Order these from broadest to narrowest - deep learning, artificial intelligence, generative AI, machine learning.

Artificial intelligence, then machine learning, then deep learning, then generative AI.

Each is a subset of the one before it.

What makes deep learning different from other machine learning?

It uses neural networks with many layers, which learn complex patterns in unstructured data such as images, text and sound.

Is every AI system a machine learning system?

No. AI is the umbrella term; machine learning is one branch of it.

A rule-based expert system is AI without machine learning.

A model reads thousands of animal photos and then draws new animals that were not in its training set. Is it discriminative or generative?

Generative. It learned the distribution of the data and creates new samples; a discriminative model would only label photos.

Which AI branch would a company use to flag inappropriate images in uploaded content?

Computer vision, which extracts information from images and video.

What separates agentic AI from a chatbot that answers one prompt at a time?

An agentic system works toward a goal on its own, planning steps and taking actions (calling tools) without a prompt for each step.

A bank has past loan records labeled "repaid" or "defaulted" and wants to predict the label for new applicants. Which learning type and problem type fit?

Supervised learning, binary classification. The data has labels and the target is one of two classes.

A retailer wants to predict next month's sales amount for each store from past data. Which ML problem type is it?

Regression, because the target is a continuous number.

A streaming company wants to group subscribers by similar viewing habits, with no predefined groups. Which ML technique fits?

Clustering, an unsupervised learning technique that finds groupings in unlabeled data.

A robot learns to walk by trying moves and getting a reward when it stays upright. Which learning type is this?

Reinforcement learning. An agent learns from rewards for its actions in an environment.

Sorting support emails into one of five topics is which classification type?

Multiclass classification. Binary classification has exactly two classes.

Do foundation models need humans to label their pre-training data?

No. They pre-train on broad unlabeled data by predicting the next item in a sequence (self-supervised).

Labeled data comes in later, for fine-tuning.

A company has no labels and wants to flag unusual credit card transactions. Which ML approach fits?

Unsupervised anomaly detection, which finds rare items that differ from the rest of the data.

A team has a few labeled medical images and many unlabeled ones. Which learning approach uses both?

Semi-supervised learning. The model trains on the labeled examples, labels the unlabeled data (pseudo-labeling), and keeps training on both.

A company wants to forecast hourly electricity demand from three years of meter readings. What type of data is this?

Time series data, measurements ordered by time.

Customer emails, product photos and call recordings are which kind of data?

Unstructured data. It has no predefined schema, unlike rows and columns in a table.

What is "ground truth" in machine learning?

A correctly labeled dataset used as the reference to train and assess a model.

A model trained on carelessly labeled images performs badly. Why?

A supervised model learns from its labels, so it can be no more accurate than its ground truth.

A spreadsheet of customers with age, income and plan columns is which data type?

Structured, tabular data.

Which neural network type is best known for image classification?

A convolutional neural network (CNN).

Which neural network type processes sequences one step at a time, for example in time-series forecasting?

A recurrent neural network (RNN), such as the one behind SageMaker AI DeepAR. Transformers later replaced RNNs for most language tasks.

Which algorithm groups customers into a chosen number of segments by similarity, with no labels?

K-means clustering.

A model must predict whether a customer will churn (yes or no) and give a probability. Which classic algorithm fits?

Logistic regression, a classification algorithm that outputs probabilities.

Which classic algorithm predicts a continuous value, such as price, from a weighted sum of features?

Linear regression (for example, SageMaker AI Linear Learner).

What is the difference between multi-class and multi-label classification?

Multi-class assigns exactly one class out of several.

Multi-label assigns every label that applies, so one item can get several.

A security team must find and locate every person in thousands of photos. Which computer vision task is this?

Object detection, which locates and classifies objects within an image.

Is logistic regression a regression algorithm for predicting numbers?

No. Despite the name, logistic regression is used for classification.

You finished every topic of AIF-C01!

Your badge turned gold. Next step: book the exam on AWS. Once you pass, add yourself to Who passed.

Share

Share that you finished this certification

I finished every topic of AWS Certified AI Practitioner (AIF-C01) on Easy Review. Come study too!