Fundamentals of AI and ML

Artificial Intelligence (AI) is the broad field of building systems that perform tasks normally requiring human intelligence. Machine Learning (ML) is a subset of AI where systems learn patterns from data rather than following explicit rules. Deep Learning is a subset of ML using multi-layer neural networks, well-suited to unstructured data like images and text.

A model is the trained artifact; an algorithm is the method used to train it (e.g., gradient descent). Training is fitting a model to data; inference is using the trained model to make predictions on new data — the exam frequently tests whether you can tell training-time concepts from inference-time concepts apart.

• Supervised learning — trains on labeled data (input-output pairs); used for classification and regression. Most production business use cases fall here.
• Unsupervised learning — finds structure in unlabeled data (clustering, dimensionality reduction, anomaly detection). No "correct answer" is provided during training.
• Reinforcement learning — an agent learns by taking actions in an environment and receiving rewards or penalties; used for sequential decision-making (robotics, game-playing, some recommendation systems).

Exam tip: if a question describes a reward signal and an agent taking actions over time, it is RL — not supervised learning, even if a table of outcomes is described.

Recognize the four data shapes the exam references: tabular (rows/columns, structured — spreadsheets, databases), text (natural language, unstructured), image (pixels, unstructured), and time-series (sequential, timestamped observations like sensor readings or stock prices).

Data also splits into labeled (has the ground-truth answer attached, required for supervised learning) and unlabeled (no ground truth, used in unsupervised learning or as a starting point for labeling pipelines).

Six stages appear repeatedly on the exam, in this order: 1. Data collection → 2. Data preparation/feature engineering → 3. Model training → 4. Model evaluation → 5. Deployment → 6. Monitoring.

Monitoring is not a one-time step — production models degrade over time as real-world data drifts away from training data (model drift), which is why continuous monitoring and periodic retraining are part of the lifecycle, not an afterthought.

Bias in an ML context means a systematic skew in predictions, usually traceable to unrepresentative training data. Fairness is the practice of measuring and mitigating that skew across protected groups (e.g., checking whether a hiring model performs equally well across demographics).

Separately, underfitting means a model is too simple to capture patterns in the data (high error on both training and test data); overfitting means a model memorized training data and fails to generalize (low training error, high test error). The exam tests whether you can pick the right fix — more data or a simpler model for overfitting; a more complex model or more features for underfitting.

A Large Language Model (LLM) is a deep learning model trained on massive text corpora to predict and generate human-like text. Generative AI is broader than LLMs — it also covers image, audio, and code generation — but LLMs are the dominant example on this exam.

Know this distinction cold: traditional/discriminative ML predicts a label or number from input; generative AI produces new content (text, images) resembling its training distribution. This framing recurs across every domain of the exam, not just this one.