AI & ML Foundations — AWS AIF-C01

Artificial intelligence is the discipline of building systems that perform tasks requiring human-like reasoning, perception, or decision-making — from image recognition to language understanding to content generation. AWS exposes AI through purpose-built services such as Amazon Rekognition for vision, Amazon Comprehend for text analysis, and Amazon Bedrock for foundation model access, letting practitioners use AI without managing research infrastructure. A common exam trap conflates AI with machine learning: rule-based expert systems are AI but not ML, because they use hand-coded logic rather than learned patterns. A system using predefined rules with no training data is AI, not ML.

Machine learning is a subset of artificial intelligence in which algorithms detect statistical patterns in training data and apply them to new inputs, without a developer coding explicit rules for every decision. On the AIF-C01 exam, ML underpins nearly every AWS AI service: Amazon SageMaker provides the managed environment for building, training, and deploying custom models, while pre-trained services like Amazon Rekognition (vision) and Amazon Comprehend (NLP) expose ML through APIs. Unlike rule-based systems, ML generalizes from examples but needs sufficient, representative data. Watch for questions on choosing a custom SageMaker model versus a ready-made AI service.

Deep learning is a subset of machine learning that stacks multiple hidden layers of artificial neurons, letting models discover hierarchical representations — edges become shapes, shapes become objects — without manual feature engineering. That automatic feature extraction makes it practical for images, audio, and natural language at scale. On the AIF-C01 exam, remember that traditional machine learning needs hand-crafted features and suits structured tabular data, while deep learning thrives on unstructured data but demands far more training data and compute. AWS delivers deep learning models through Amazon Bedrock (managed foundation models) and Amazon SageMaker (custom training and hosting).

A neural network is a computational model built from layers of interconnected nodes that transform inputs into outputs by adjusting numerical weights through training. Data flows from an input layer through one or more hidden layers to an output layer; the number and depth of those hidden layers determine the network’s capacity to learn complex patterns. The key exam distinction is between a shallow network (one hidden layer) and a deep neural network (many hidden layers) — deep learning is simply this increased depth applied at scale. Amazon SageMaker is the primary AWS service for training custom neural networks, while Amazon Bedrock exposes pre-trained foundation models that are themselves large neural networks, without requiring users to train the underlying architecture.

Supervised learning trains a model on data where every input is paired with a correct output label, so it learns a mapping it can apply to new, unlabeled inputs. The two core task types are classification (predicting a discrete category, like spam or not spam) and regression (predicting a continuous value, like a price). Amazon SageMaker supports this full workflow through built-in algorithms and custom scripts. The key exam distinction: supervised learning requires human-annotated labels before training, whereas unsupervised learning finds structure in unlabeled data. Scenarios mentioning “known outputs” or “historical examples with correct answers” signal supervised learning.

Unsupervised learning trains models on data with no human-provided labels, letting the algorithm discover hidden patterns on its own. Common techniques include clustering (grouping similar points, as in customer segmentation), anomaly detection (flagging outliers), and dimensionality reduction. Amazon SageMaker offers unsupervised built-in algorithms such as K-Means and Random Cut Forest. The key exam distinction: unsupervised learning has no ground-truth target variable to optimize against, unlike supervised learning where a labeled outcome drives training. That absence makes evaluation harder, since there is no straightforward accuracy metric.

Reinforcement learning (RL) is a paradigm where an agent observes its environment, takes actions, and receives scalar feedback — rewards for desirable outcomes, penalties for undesirable ones — iterating until it learns a policy that maximizes cumulative reward. Unlike supervised learning, it needs no labeled input-output pairs; the agent explores to discover which actions yield the best long-term returns. On AIF-C01, the key application is reinforcement learning from human feedback (RLHF), used to align large language models with human preferences after pre-training. Expect conceptual questions about RLHF rather than service-specific implementation details.

Inference is the process of running new, unseen input through an already-trained model to produce a prediction, classification, or generated response. Unlike training — which adjusts model weights over large datasets and is compute-intensive — inference applies fixed weights, making it far faster per request but the primary source of recurring operational cost at scale. A key exam distinction is real-time inference (low-latency, synchronous, suited for interactive applications) versus batch inference (asynchronous processing of large datasets, lower cost per item). Amazon SageMaker supports both through endpoints and batch transform, while Amazon Bedrock exposes foundation-model inference via a managed API without provisioning infrastructure.

Training data is the labeled or unlabeled dataset a model learns from, and its quality and representativeness set the ceiling on model performance. In Amazon SageMaker, practitioners typically supply training data from Amazon S3, splitting it into training, validation, and test sets to control learning and unbiased evaluation. The exam stresses the training-versus-test distinction: training data shapes the model’s weights, while test data must stay entirely unseen until final evaluation to yield a valid accuracy estimate. Leaking test data into training inflates reported performance, a common scenario trap on AIF-C01.

Bias is systematic error that makes a model produce skewed or unfair outputs, often because training data over- or under-represents certain groups. On AIF-C01 it has two senses: statistical bias (error from oversimplified algorithm assumptions) and societal bias (unfair outcomes for protected groups). Amazon SageMaker Clarify measures both pre-training data bias and post-training model bias, reporting metrics like class imbalance and disparate impact. The exam also tests bias versus variance: high bias underfits (too simple), high variance overfits (too complex). Using more complex models to cut bias can raise variance, so the two must be balanced.