Color	Field	What it means
Teal	Data Science	The broadest field — using data, statistics, and tools to find patterns and solve problems
Purple	Analytics	Turning data into insights (what happened, why, what might happen next)
Blue	AI	Machines doing tasks that usually need human intelligence
Green	ML	AI that learns from data instead of following only fixed rules
Orange	DL	ML using deep neural networks (many layers)
Magenta	Generative AI	DL systems that create new content — text, images, code, audio

Real-world examples

• Analytics: “Sales dropped 12% in Q3” (dashboards, reports)
• ML: “This customer is likely to churn” (prediction from data)
• DL: “This MRI scan shows a tumor” (image recognition)
• Generative AI: “Write me a summary of this report” (creating new content)

Features and Labels

Core idea

Role	What it is	Example
Feature	Input used to predict or find patterns	age, income, click_count
Label	Output you want to predict or explain	churned, house_price

Main terminology

Concept	Data Science / ML	Statistics	Other popular names
Feature	Feature	Predictor, Independent variable (IV), Covariate, Regressor	Input, Attribute, Variable, Column, Dimension, Field
Label	Label, Target	Response, Dependent variable (DV), Outcome	Output, Ground truth, Y, Class, Category

Features = what you know (inputs). Labels = what you want to predict (outputs). Different fields use different words for the same roles.

Classification

Supervised learning where the label is a category (discrete class). The model learns to assign new inputs to one of a fixed set of classes.

Example — Email spam filter: Features include email text, sender domain, and number of links. The label is spam or not_spam.

---

Regression

Supervised learning where the label is a number (continuous value). The model learns to predict a quantity on a numeric scale.

Example — House price prediction: Features include square footage, bedrooms, and zip code. The label is the sale price (e.g. $425,000).

---

Unsupervised

Learning from features only — no labels. The model finds hidden structure such as groups, compressed representations, or outliers.

Example — Customer segmentation: Features include annual spend and visit frequency. The model groups customers into clusters like “loyal” or “at-risk” without predefined labels.

---

Reinforcement

An agent learns by taking actions and receiving reward signals through trial and error — not from a fixed labeled dataset.

Example — Chess AI: State is the board position. The agent picks a move and gets +1 for a win, −1 for a loss. It learns a strong move policy over time.

---

Semi-supervised

Combines a small labeled set with a large unlabeled pool when labels are expensive to collect.

Example — Medical imaging: 200 labeled X-rays (tumor / no_tumor) plus 10,000 unlabeled scans to classify new images.

---

Self-supervised

Labels are created from the data itself to learn useful representations before a downstream task.

Example — BERT pre-training: “The cat sat on the ___” → predict mat → later fine-tuned for sentiment analysis.

---

Transfer learning

Reuses a model trained on one task and adapts it to a new, related task with limited data.

Example — Product classifier: 50 labeled product photos, starting from an ImageNet-trained model, to classify new products accurately.

Parametric vs Non-Parametric Methods in Machine Learning

Parametric methods assume the relationship between inputs and outputs has a fixed shape before you train. You only estimate a set number of parameters — for example, the slope and intercept in linear regression. The model stays the same size whether you have 100 data points or 100,000.

Non-parametric methods do not lock in one global shape. They learn patterns more flexibly from the data, and their complexity can grow as you add more data. They often rely on ideas like “similar examples should give similar answers” rather than “the relationship is always a straight line.”

Simple analogy

• Parametric: You decide the recipe is “a straight line” and only learn how steep it is and where it starts.
• Non-parametric: You don’t decide the shape upfront. You let the data guide a more flexible fit, like connecting nearby points or splitting the space into regions.

Examples

• Parametric: linear regression, logistic regression, Naive Bayes, Gaussian mixture models, neural networks with a fixed architecture.
• Non-parametric: k-nearest neighbors, decision trees, kernel methods, Gaussian processes.

Main tradeoff

• Parametric: simpler, faster, easier to interpret — but wrong if your assumed shape doesn’t match reality.
• Non-parametric: more flexible, better for complex patterns — but usually needs more data and can overfit more easily.