Fundamentals of AI and ML
What you will learn
In this domain, you learn the relationship between artificial intelligence (AI), machine learning (ML), and deep learning (DL), the differences between supervised, unsupervised, and reinforcement learning, and the general lifecycle of running an ML project. It is a foundational domain that accounts for 20% of the AIF-C01 exam and forms the basis for everything that follows.
Key points
- The containment relationship AI ⊃ ML ⊃ DL - AI is the broadest concept, ML is one approach within AI, and DL is an ML approach using neural networks
- Generative AI and agentic AI - generative AI is an application of deep learning that creates new content with foundation models, and agentic AI is AI that autonomously works through multiple steps and tools toward a goal
- Traditional ML or a foundation model - the stricter the regulatory requirements, the need for explanations, and the operational constraints (response time and cost), the better traditional ML fits; the wider the range of inputs a use case handles, the better a foundation model fits
- Supervised learning - learns from pairs of inputs and correct answers (labels). It is divided into classification (categorical) and regression (numerical)
- Unsupervised learning - discovers the structure of data without correct labels. Representative examples are clustering, dimensionality reduction, and anomaly detection
- Reinforcement learning - an agent interacts with an environment and learns actions that maximize reward
- The ML lifecycle - the repeating cycle of data collection, preprocessing, training, evaluation, deployment, and monitoring
- AWS services for each stage of the pipeline - SageMaker AI for developing and operating your own models, Bedrock for building in foundation models, Kiro for writing code, and Amazon Quick for business use (the exam guide 1.1 (English version) also lists Amazon Q as an example)
- Overfitting - the phenomenon in which performance fits the training data but drops on test data. Care is needed to preserve generalization performance
- Inference - the process of giving a trained model new input to obtain predictions. There are four methods: real-time, batch, asynchronous, and serverless
Terms and concepts
The relationship between AI, machine learning, and deep learning
AI is a broad concept that refers to technologies that mimic intelligent human behavior in general. ML is an AI approach that "automatically learns rules from data," and DL is an ML approach that "uses multilayer neural networks." Most generative AI is an application of DL.
Where generative AI and agentic AI fit
Generative AI is the field that creates new content such as text, images, and code by using foundation models built with deep learning. Agentic AI is AI that makes a plan toward a given goal and autonomously works through multiple steps and tool calls. Whereas traditional AI responds once each time it receives an instruction, agentic AI repeats judgment and action, using the reasoning of a foundation model, until it reaches the goal.
Traditional ML or a foundation model
For reviews where the basis of a decision must be explained to regulators or customers, a traditional ML model with a simple, easy-to-explain mechanism fits well. A small traditional model also fits when there are operational constraints, such as strict limits on response time or cost per request. Foundation models fit situations such as text summarization and conversation, where the use case handles a wide range of inputs and it is hard to prepare training data for each use. If the processing can be written as fixed rules, you do not use ML in the first place.
Supervised learning and unsupervised learning
Supervised learning learns from labeled data and is used for image classification and sales forecasting. Unsupervised learning finds structure in unlabeled data and is used for customer segmentation and anomaly detection. Semi-supervised learning is an intermediate between the two.
Reinforcement learning
A technique in which an agent (the learner) selects actions toward an environment and learns a strategy that maximizes the reward obtained as a result. It is used in game AI, robot control, recommendation optimization, and RLHF (reinforcement learning from human feedback) for generative AI.
Overfitting and generalization
Overfitting is the phenomenon in which a model fits the training data too closely and performs poorly on unseen data. It is prevented by increasing the amount of data, regularization, dropout, and cross-validation. Conversely, underfitting is a state in which the model is too simple to learn even the training data. The degree of generalization is measured on test data that was not used for training, and for classification models you look not only at accuracy but also at precision (the share of predicted positives that were actually positive), recall (the share of actual positives that were found), and the F1 score, the harmonic mean of the two.
AWS services for each stage of the AI/ML pipeline
At the stage where you carry out everything yourself, from data preparation through model training, evaluation, deployment, and monitoring, you use Amazon SageMaker AI. At the stage where you call foundation models through an API without training and build them into applications, Amazon Bedrock is central. Kiro handles the stage of writing the code for those applications, and Amazon Quick handles the stage of using the finished AI in business work. The exam guide 1.1 (English version) also lists Amazon Q as an example (names and categories as of October 2026).
→ Amazon Bedrock Glossary / → Spec-Driven Development with Kiro / → Amazon Quick in Practice
Methods of inference
Real-time inference returns a prediction immediately for each request and is used for use cases that need low latency (e.g., a chatbot). Batch inference processes a body of data all at once (e.g., predicting purchases for all customers overnight) and does not need an endpoint that runs all the time. Asynchronous inference queues requests and processes them in order; it suits large inputs or long processing times, and it can scale instances down to zero while there are no requests. Serverless inference is a pay-per-use method that you can use without choosing instance types or scaling settings, and it scales automatically with the volume of calls. It suits use cases that have idle time between calls and can tolerate the startup delay (cold start).