Artificial intelligence, machine learning and deep learning name three different things, and the three names are used as one in most conversations about a product. Artificial intelligence is the widest of the three and dates to the 1950s. Machine learning sits inside it, and deep learning sits inside machine learning. The AWS AI Practitioner exam opens with this vocabulary because a team that cannot place a system in that hierarchy cannot say what the system will do when it meets an input nobody anticipated.
The sections below take the three terms in the order they nest, then the four words the syllabus uses for the mechanics, then bias, fairness and fit, and end with the AWS services the exam attaches to each term.
How the three terms nest
Artificial intelligence is the outer field. Any technique that makes a machine behave in a way people would call intelligent belongs to it, including techniques that do no learning at all. A rules engine that reads a written credit policy and decides eligibility is artificial intelligence.
Machine learning is the part of that field where behaviour is learned from data. Nobody writes the rule down. The rule is inferred from examples, which is what makes machine learning useful on problems too messy to specify and hard to explain once it works.
Deep learning is the part of machine learning that uses neural networks of many layers. A deep network learns its own features. Classical machine learning needs a person to decide that the useful signal in a photograph is edge density, and a deep network works that out for itself given enough examples. That is why deep learning took over computer vision, speech recognition and natural language processing, and why it needs far more data and compute than the classical methods.
The nesting is what the vocabulary questions test. A system can be artificial intelligence with no machine learning in it at all, and every deep learning system is a machine learning system.
Algorithm, model, training and inference
The syllabus separates four words that teams use interchangeably.
An algorithm is the procedure that learns. A model is the artefact that procedure produced. Training is running the algorithm over data to produce the model. Inference is using the finished model to produce an answer for an input it has never seen.
The line between training and inference decides most of a system's cost and most of its architecture. Training is expensive, occasional and done offline. Inference costs little for a single call, runs constantly and sits in the path of a user waiting for an answer.
Bias, fairness and fit
Three more words carry a precise meaning in the syllabus and a looser one in conversation.
Bias has two meanings that are easy to confuse. Statistically, bias is error from a model too simple to capture the pattern. Socially, bias is a system producing systematically different outcomes for different groups. Both appear in the syllabus, and the social meaning is the one governance exists to address.
Fairness is whether those different outcomes can be justified. Fairness has no single measurement, and the published definitions of it conflict with one another mathematically, so it stays a judgement and never becomes a threshold.
Fit describes how closely a model matches its training data. A model that has learned the noise along with the signal is overfitted and does badly on anything new. A model too simple to capture the pattern is underfitted and does badly everywhere.
Where AWS puts each of these
The exam is vendor specific, so the vocabulary arrives attached to services. Amazon SageMaker is the platform for building and training a model the organisation owns. Amazon Bedrock serves foundation models the organisation did not train. Amazon Rekognition, Amazon Transcribe, Amazon Comprehend, Amazon Polly, Amazon Translate and Amazon Lex are prebuilt services for vision, speech, text and conversation, and they take no training data from the customer at all.
The prebuilt services are the ones candidates underestimate. A large part of the exam is recognising a problem a managed service already solves, so that building and training a model would be wasted work.