AI vs ML vs Deep Learning: Core AIF-C01 Concepts and Terminology

These five terms are not competing labels; they nest inside one another, and the exam expects you to know the ordering. Artificial intelligence (AI) is the widest circle: any technique that lets a computer mimic human intelligence, including old-fashioned rule-based systems that follow hand-written logic. Machine learning (ML) sits inside AI. It is the family of methods that learn patterns from data rather than being explicitly programmed with rules. Deep learning sits inside ML. It uses neural networks with many stacked layers to learn from large amounts of data, and it powers most modern breakthroughs in images and language. Generative AI (GenAI) is built on deep learning. It uses very large models (foundation models and large language models) to generate new content such as text, images, or code.

Agentic AI is the newest term. An agentic system uses a generative model as a reasoning engine, then adds memory and the ability to call tools so it can plan and take multi-step actions toward a goal, not just answer a single prompt. Think of it as GenAI that acts. The table below fixes the hierarchy in one view, and the next section separates the terms that are most often confused with one another.

Generative AI creates new content, while traditional predictive ML outputs a label, score, or number about input that already exists. A credit model that rates a loan applicant's risk of default, a churn model that scores each customer, and a forecast of next week's sales are all predictive: nothing new is written or drawn. A model that drafts a marketing email, summarizes a report, or creates an image is generative. Both are machine learning; the difference is the kind of output.

Generating content is not, by itself, agentic. A model that writes one description, one email, or one answer per request is generative AI, however good the output. A system becomes agentic when the model itself decides what to do next: it breaks a goal into steps, chooses which tool or system to call, reads the result, and picks the following step from what it learned, often keeping memory of earlier results along the way. The test is who chooses the path. If a developer has hard-coded the sequence, so every request follows the same path whatever the model says, the application is a fixed workflow that uses generative AI, even though it calls APIs and the model writes part of the output. If the model plans and selects the steps at run time, it is agentic AI.

Two other terms are frequently placed next to agentic AI. Retrieval-augmented generation (RAG) grounds a model's answer in documents retrieved from a knowledge source at question time; it improves what the model says but does not plan or take actions (Domain 3 covers it in depth). The "agent" in reinforcement learning is a different idea: a learner trained by rewards and penalties through trial and error in an environment. An agentic AI system is usually built on a pre-trained foundation model and is not trained by rewards while it does its job, so the shared word "agent" should not lead you to treat the two as the same thing.

Four foundational words appear in almost every AIF-C01 question about how machine learning works, and distractors love to swap them. An algorithm is the method or procedure used to learn from data, for example a decision-tree or gradient-descent procedure. A model is the result you get after running that algorithm over data: the trained artifact that stores the learned patterns and can make predictions. In short, the algorithm is the recipe and the model is the cake. Because the learned patterns come from the data, one algorithm can yield many models: train it on last year's records and again on this year's and you have two distinct models. Hyperparameters are the settings you choose before training starts, such as the maximum depth of a decision tree or how many passes to make over the data; they shape how the algorithm learns but are inputs to training, not something training produces. (Tuning them is part of the ML lifecycle in Task 1.3.)

Training is the phase where the algorithm processes example data and adjusts the model until it captures useful patterns. It is compute-heavy and usually happens before the model is used. Inferencing (also called inference) is the phase where the finished model is given new, unseen input and produces a prediction or output. Training builds the model; inferencing uses it. Inferencing leaves the model as training made it: the input arrives without an answer attached, the model processes it and returns an output, and nothing is learned from that request unless you later retrain on new data. A helpful test for the exam: if the sentence describes the system learning from labeled or historical examples, it is training. If it describes the system answering a live request or scoring new records, it is inferencing. The costs behave differently too: training is a periodic expense paid when you build or rebuild the model, while inferencing is paid per use, so a popular application's inference bill rises with its traffic. That is why the two are budgeted and scaled separately.

These terms name the technologies and problem areas the exam references repeatedly. A neural network is a model loosely inspired by the brain: layers of connected nodes that pass signals forward, with each connection weighted so the network can learn complex relationships. A neural network with many hidden layers is what makes something deep learning. What a network learns is stored as millions or billions of numeric weights spread across those layers, not as rules a person can read.

Deep learning versus traditional ML

Deep learning learns useful features directly from raw inputs, while traditional ML methods such as decision trees and gradient-boosted trees usually depend on people engineering input features first. In speech recognition, for instance, the early hidden layers learn short sound patterns and later layers combine them into syllables and words, so engineers do not have to design the acoustic measurements first. Tree-based methods, by contrast, split on one input feature at a time, which suits well-defined columns but is a poor match for raw, high-dimensional signals such as a waveform or an image. That power has costs. Deep networks typically need far more training data and compute, and because their knowledge is spread across weights they are harder to interpret. A single decision tree stores what it learns as branching if-then splits you can read and explain; ensembles such as gradient-boosted trees combine hundreds of trees, so they are less readable than one tree but still usually easier to explain than a deep network.

The rule of thumb: when data is limited and already arranged in columns, traditional ML usually trains faster, performs as well or better, and is easier to justify; when data is plentiful and raw (images, audio, free text), deep learning usually wins.

Computer vision is the field concerned with helping machines interpret images and video, such as detecting objects, reading faces, or spotting defects on a production line. Natural language processing (NLP) is the field concerned with understanding and working with human language, covering tasks like sentiment analysis, translation, and entity extraction. Speech recognition, which converts spoken audio into text, is closely related and often grouped with NLP applications.

A large language model (LLM) is a very large neural network trained on huge volumes of text so it can predict and generate language. It generates text one token (roughly a word or part of a word) at a time by predicting what should come next. LLMs are the engines behind most generative AI text applications and behind agentic systems. Two neighbouring model types are easy to mix up with it: an embedding model turns text into numeric vectors that capture meaning, which powers semantic search and similarity matching, but it does not generate replies; a diffusion model generates images by gradually turning random noise into a picture. Domain 2 covers both in more depth. For Task 1.1 you only need to recognize what each term means and which problem it addresses; the deeper mechanics of foundation models and tokens belong to the generative AI domain, not here. Being able to say "computer vision handles images, NLP handles language, an LLM generates text" is enough to answer definition questions correctly.

This cluster of terms describes how well a model behaves, and it reappears in the responsible-AI domain, so learning it now pays off twice. Bias, in the fairness sense, is a systematic skew in a model's outputs, often caused by unrepresentative training data, that produces unfair results for some groups of people. Fairness is the goal of treating individuals and groups equitably so the model does not disadvantage anyone based on attributes like age, gender, or ethnicity. Bias is the problem; fairness is the objective you pursue by detecting and reducing that bias. A related but different term is explainability: the ability to understand and describe why a model reached a particular decision. Explainability helps you investigate a suspected bias, but a lack of it is not the same thing as bias.

Fit describes how well a model matches the underlying pattern in data, and the exam tests two failure modes. Overfitting happens when a model learns the training data too closely, including its noise, so it performs well on examples it has seen but poorly on new data; it has essentially memorized instead of generalized. Underfitting is the opposite: the model is too simple to capture the real pattern, so it performs poorly even on the training data. The sweet spot, good fit, generalizes well to unseen data. A quick discriminator: great on training data but bad on new data means overfitting, while bad on both means underfitting.

Statisticians also use "bias" in a second, technical sense, paired with variance. Variance is how much a model's predictions change when it is trained on a different sample of data; a high-variance model is overly sensitive to its training set, which is the root of overfitting. Statistical bias here means error from overly simple assumptions, which is the root of underfitting. Keep the senses apart: when a scenario describes a model treating a group of people unequally, the terms that apply are bias (the unfair skew) and fairness (the goal), not variance or overfitting.

The AIF-C01 exam expects you to tell three learning types apart by the kind of data they use and the problem they solve. Supervised learning trains on labeled data, meaning each example already carries the correct answer, and the model learns to map inputs to those answers. A spam filter trained on emails each marked spam or not-spam is supervised learning. Unsupervised learning trains on unlabeled data and finds hidden structure on its own, with no correct answers provided. Grouping customers into segments by their purchasing behavior, with no predefined categories, is unsupervised learning. Reinforcement learning has no fixed dataset of answers at all; an agent interacts with an environment and learns by trial and error, receiving rewards for good actions and penalties for bad ones. Training a robot to walk or a program to play a game are classic examples.

The single most useful clue is the word "labeled": if a scenario says the data is labeled or has known outcomes, think supervised; if it says items must be grouped with no labels, think unsupervised. The setting does not change the learning type. A driving model trained on recorded journeys, each moment tagged with the steering an expert driver applied, is supervised learning, because it learns from labeled examples; it would be reinforcement learning only if it learned by driving in a simulator and receiving rewards for good outcomes.

A fourth term appears beside these three. Self-supervised learning creates its own labels from the raw data, for example by hiding the next word in a sentence and learning to predict it. This is how large language models are pre-trained on huge volumes of unlabeled text. It differs from supervised learning, where the labels are supplied from outside the data (by people or an existing record), such as an email marked spam or a transaction confirmed as fraud.

Models are only as good as the data they learn from, and the exam asks you to recognize common data categories. The broadest split is structured versus unstructured. Structured data is organized into a predictable format, typically rows and columns, so it is easy to store in tables and query. Two important structured kinds are tabular data (spreadsheet-style rows and columns, such as customer records) and time-series data (values recorded in time order, such as daily temperature or hourly sales). Unstructured data has no fixed schema and includes image, text, and audio, which make up the large majority of data organizations hold.

A second, separate distinction is labeled versus unlabeled data, which ties directly to the learning types above. Labeled data comes with the correct answer attached to each example, for instance a photo tagged "cat," and it is what supervised learning requires. Unlabeled data has no such answers and is what unsupervised learning works with. Labeling is often expensive and time-consuming because it usually needs humans, which is a practical reason teams sometimes prefer unsupervised approaches. Keep the two axes distinct on the exam: structured-versus-unstructured is about format, while labeled-versus-unlabeled is about whether answers are provided.

Once a model exists, there are several ways to run inference, and Task 1.1 asks you to match the style to the workload. Batch inferencing processes a large volume of data all at once, offline, when you do not need an immediate answer, for example scoring every customer overnight to update a churn report. Real-time inferencing returns a prediction immediately for a single request and is used when low latency matters, such as approving or declining a card payment at checkout. Choose real-time when a user or system is waiting on the answer right now.

Asynchronous inferencing suits requests with large payloads or long processing times: the request is queued, processed when resources are ready, and the result is delivered when it is done, so the caller is not blocked while waiting. It is a middle ground between batch and real-time, useful for tasks like processing a large document or video. Serverless inferencing removes the need to manage or pre-provision servers; capacity scales automatically with demand, including scaling down when idle, which fits intermittent or unpredictable traffic where you would rather not pay for idle infrastructure. Serverless is still a synchronous request-response pattern for small, quick requests: the caller sends a request and waits for the answer, as with real-time inferencing, and it is not a queue for large or long jobs. The trade-off for scaling to zero is a cold start: after an idle period, the first request waits while compute is brought up, so it is slower than the requests that follow. A provisioned real-time endpoint avoids that delay but keeps paying for capacity while it sits idle.

The pattern to remember: many records at once and no rush means batch, an instant single answer under steady load means real-time, large or slow jobs mean asynchronous, and quick requests with spiky or occasional traffic, where idle cost matters more than an occasional cold start, mean serverless.

A retailer wants to reduce fraud and understand its customers, and its project touches nearly every Task 1.1 term. Its data science team starts with structured, tabular transaction records. Because past transactions are labeled as fraudulent or legitimate, they use supervised learning: an algorithm processes the labeled history during training and produces a model. To catch fraud the instant a purchase is made, the model is deployed for real-time inferencing at checkout, while a nightly job re-scores dormant accounts using batch inferencing.

Separately, the team has millions of unlabeled product-review texts. With no answers attached, they apply unsupervised learning to group reviews into themes, an NLP task. A computer vision model checks product photos for quality, and a generative AI assistant powered by a large language model drafts replies to customers; when that assistant is allowed to look up an order and issue a refund on its own, it becomes agentic AI. Along the way the team watches for overfitting, so the fraud model does not just memorize old cases, and audits for bias to keep decisions fair across customer groups. Every one of these words is a definition the exam can ask you to recognize.