Roadmap › Part I · The AI Landscape › Module 0 · Orientation

Lesson 03 · Foundations

AI vs ML vs Deep Learning vs Data Science

Four words are used almost interchangeably in job postings, headlines, and conference talks — and the confusion is not harmless. Getting the relationships exactly right tells you what each field does, what it can't, and precisely where this course sits.

Reading: ~22 min Prerequisites: Lessons 1–2 Code: none — this is the map, sharpened

Why a whole lesson on four words

Lesson 1 gave you a rough containment sketch (AI ⊃ ML ⊃ NN ⊃ DL). That was enough to start. But these four terms are the ones you'll be asked about in interviews, the ones marketing departments blur on purpose, and the ones that decide whether a project needs a laptop or a GPU cluster. By the end you'll be able to take any real system or job description and say, without hesitation, which boxes it lives in and why — including the four adjacent terms (neural networks, generative AI, LLMs, data science) that people constantly mis-file.

What this lesson delivers

The precise containment relationship between AI, ML, Neural Networks, and Deep Learning — and why each nesting is real, not marketing.
Why Data Science is not in that nesting at all, but an overlapping discipline.
Where Generative AI and LLMs actually sit (spoiler: two levels below where we stop).
A concrete, side-by-side comparison you can apply to any system.
A decision rule for choosing classical ML vs Deep Learning — the practical payoff.

1 · The nesting: four Russian dolls

The single most useful mental image is a set of nested circles — each term is a strict subset of the one before it. Let's build it from the outside in, and for each layer state the defining move that separates it from its parent.

Artificial Intelligence — the outermost circle. Any technique that makes a machine act intelligently toward a goal: search, hand-written logic, statistical learning, or bio-inspired methods. Defining move: goal-directed behavior, by any means — including means with no learning at all.

Machine Learning ⊂ AI — the subset where behavior is learned from data rather than explicitly programmed. Defining move: the rules come from examples, not from a person writing them down. This is our course.

Neural Networks ⊂ ML — one family of ML models, loosely inspired by interconnected brain neurons: layers of simple units with tunable weights. Defining move: the model is a network of weighted connections, not a tree, a line, or a distance rule.

Deep Learning ⊂ Neural Networks — neural networks with many hidden layers ("deep"), which lets them learn features automatically. Defining move: depth — enough layers that the network discovers its own representations instead of being handed them. Deliberately outside our scope.

Read the nesting as a chain of "how"

Each layer answers "how does the layer above do its job, in this particular case?" How does this AI act intelligently? By learning from data → it's ML. How does this ML model represent the problem? As a network of neurons → it's a neural network. How deep is that network? Many layers → it's Deep Learning. Every step inward is a more specific commitment, and every step inward is a strictly smaller set of systems.

The nesting also means the arrows only point one way. Every Deep Learning system is Machine Learning; not every Machine Learning system is Deep Learning. A GPS route-finder is AI but not ML (it searches, it doesn't learn). A decision tree is ML but not a neural network. This directionality is exactly what people get wrong when they call every clever program "AI" or every ML model "deep learning."

The one-line test for each boundary

Is it AI? Does it act toward a goal on a task we'd call intelligent? (Even hand-coded rules count.)
Is it ML? Would its behavior change if you fed it different training data — without you rewriting code?
Is it a neural network? Is the model literally layers of weighted units?
Is it Deep Learning? Does that network have many layers that learn features on their own?

2 · The boundary that matters most: classical ML vs Deep Learning

The nesting is tidy, but as an engineer the practically important line is between layers 2 and 4: classical (statistical) Machine Learning — where this course lives — versus Deep Learning. They are not rivals so much as tools for different regimes. Here's the honest comparison.

Dimension	Classical ML (this course)	Deep Learning
Best data type	Structured / tabular (rows & columns, spreadsheets, databases)	Unstructured (images, audio, raw text)
Features	You engineer them — pick and shape the input columns	The network learns them from raw input
Data hunger	Works well on hundreds–thousands of rows	Typically needs tens of thousands to millions
Compute	Trains on a laptop CPU in seconds–minutes	Often needs GPUs/TPUs and hours–days
Interpretability	Often high (a tree or linear model explains itself)	Usually low (a "black box" needing extra tooling)
Typical wins	Fraud, churn, pricing, risk, most business/tabular problems	Vision, speech, translation, generative text/images

Why we stop at classical ML — and why that's the right call

Deep Learning gets the headlines, but classical ML is still the correct tool for the majority of real-world problems, which are tabular, data-limited, and interpretability-sensitive. More importantly, classical ML is where every core idea — model, loss, optimization, overfitting, evaluation — appears in its simplest, most visible form. Deep Learning reuses all of them, just buried under layers. Learn them here, cleanly, and the deep stuff later becomes "more of the same, scaled up" rather than a new mystery.

The key structural difference is worth dwelling on: feature engineering. In classical ML, you decide that "price per square meter" or "days since last purchase" is the meaningful signal — the model only fits weights to features you designed. In Deep Learning, you hand the network raw pixels or characters and its many layers discover useful features themselves. That single shift — who invents the features — is the deepest reason Deep Learning needs so much more data and compute, and why it becomes a black box. Depth buys automatic feature-learning; the price is data, compute, and interpretability.

3 · Data Science: the term that is not in the nesting

Here is where most explanations go wrong. Data Science is not a doll inside the AI set. It overlaps AI/ML, but it is a distinct discipline with a different goal.

Working definition

Data Science is the interdisciplinary practice of extracting insight and value from data. It blends statistics, programming, and domain expertise, and its deliverable is often a decision, a report, or a dashboard — not necessarily a deployed model. Machine Learning is one tool in its kit, alongside data cleaning, exploratory analysis, visualization, and classical statistics.

The clarifying question is what is the deliverable?

Machine Learning's deliverable is a model — an artifact that makes predictions on new inputs. The goal is predictive performance.
Data Science's deliverable is often understanding — "why did churn rise last quarter?", "which customers matter most?". The goal is insight that guides a human decision.

These overlap heavily: a data scientist frequently trains ML models, and an ML engineer does plenty of data analysis. But the fields are not nested. Plenty of high-value data science uses no ML at all (a well-designed A/B test, a cohort analysis, a clear chart), and plenty of ML is pure engineering with little "science" (retraining a known model on fresh data in a pipeline).

	Primary goal	Typical deliverable	Relationship
AI	Make machines act intelligently	An autonomous/goal-seeking system	The umbrella field
ML	Predict well from data	A trained model	Subset of AI; tool of Data Science
Deep Learning	Learn from raw, unstructured data	A trained deep network	Subset of ML
Data Science	Extract insight to guide decisions	Reports, dashboards, decisions, sometimes models	Overlaps ML — not nested

4 · Where Generative AI and LLMs actually live

Because you'll be asked about them constantly, let's place today's headline technologies precisely on the map.

→

Generative AI — models that produce new content (text, images, audio, code) rather than only labeling or scoring existing input. It is a capability, and today it is delivered almost entirely by Deep Learning — so it sits inside box 4.

→

Large Language Models (LLMs) — ChatGPT, Claude, and kin: very large deep neural networks (the "transformer" architecture) trained on enormous text corpora. They are a specific kind of Generative AI, so they nest even deeper: LLM ⊂ Generative AI ⊂ Deep Learning ⊂ Neural Networks ⊂ ML ⊂ AI.

The scope boundary, stated plainly

LLMs and generative models are two levels below where this course stops. That is not because they're unimportant — it's because they're deep-learning applications, and you cannot understand a transformer without first understanding the ML fundamentals (loss, gradient descent, overfitting, evaluation) that this course teaches in their clean, classical form. We mention generative AI only to locate it on the map. When you finish here, you'll have the exact foundation those systems are built on.

5 · The map applied — worked classifications

Abstractions stick when you run them against real systems. For each, we tag every box it occupies.

System	AI?	ML?	Deep Learning?	Data Science?	Why
GPS shortest-route finder	Yes	No	No	No	Search/planning — goal-directed but learns nothing from data
Spam filter trained on your "mark as spam"	Yes	Yes	Usually no	Overlaps	Learns rules from labeled examples; classical ML suffices
Churn dashboard + cohort analysis	No*	Maybe	No	Yes	Insight for humans; may use ML but often just statistics
Face recognition on your phone	Yes	Yes	Yes	No	Deep network learning features from raw pixels
ChatGPT / Claude	Yes	Yes	Yes	No	LLM ⊂ Generative AI ⊂ Deep Learning
Credit-scoring logistic regression	Yes	Yes	No	Overlaps	Classical ML on tabular data — interpretable by design

*A pure analysis/dashboard with no autonomous decision-making is Data Science but wouldn't usually be called an "AI system" — a good reminder that the terms answer different questions.

6 · The practical payoff — a decision rule

The whole point of getting the map right is to make better choices. Here's the rule of thumb an experienced practitioner actually uses when starting a supervised problem:

Classical ML or Deep Learning? Start here

Is your data tabular/structured? → Start with classical ML. It will train in seconds, be interpretable, and is very often the final answer. Reach for Deep Learning only if it clearly underperforms.
Is your data raw images, audio, or free text, and do you have lots of it? → Deep Learning is likely the right tool (outside this course's scope, but now you know why).
Do you have little data, or must you explain every decision (regulated finance, medicine, law)? → Strongly favor classical ML — it needs less data and can be made transparent.
Not sure? → Always baseline with classical ML first. A simple model you understand beats a complex one you don't, and it tells you whether the extra machinery is even worth it.

Engineer's translation

Treat model complexity like you treat any dependency: add it only when a simpler thing has demonstrably failed. "Just use a neural network" is the ML equivalent of pulling in a giant framework to do string concatenation. The map you now hold is what lets you make that call deliberately instead of by hype.

7 · Limitations — where the neat map gets fuzzy

Honest caveats to carry forward

The boundaries are conventions, not laws. How many layers make a network "deep"? There's no official threshold. Treat the nesting as a communication tool, not a theorem.
Real systems are hybrids. A self-driving car mixes deep learning (perception), search (planning), and hand-coded rules (legal constraints). "Which box?" is often "several at once."
Marketing muddies everything. Products slap "AI" on plain if-statements and "deep learning" on shallow models. Your job is to ask the one-line boundary tests, not trust the label.
Terms drift over time. "AI" today colloquially often means "generative AI," which would have baffled a 1990s researcher. Anchor to the definitions, and note when someone is using the popular sense.

How this connects to the next lesson

Bridge to Lesson 4

You can now place any system on the map, and you've committed to box 2 — classical Machine Learning — with clear eyes about what's above and below it. In Lesson 4 — What Is Machine Learning? we stop pointing at the box from outside and step inside it: the precise definition of learning, the three types (supervised, unsupervised, reinforcement), and the end-to-end ML workflow that every remaining lesson instantiates. This is where the map becomes a machine.

Exercises & reflection

Still no coding — these sharpen the classifications you'll rely on for the rest of the course.

Exercise 1 · Tag the boxes

For each system, mark which of {AI, ML, Deep Learning, Data Science} apply, and justify in one sentence:

A recommender that suggests movies based on what similar users watched.
A quarterly sales report with charts and a trend line, built by hand in a notebook.
A voice assistant that transcribes your speech to text.
A thermostat schedule you programmed with fixed if time > 18:00 rules.

Show discussion

(1) AI + ML, overlaps Data Science; usually classical ML (collaborative filtering), not deep. (2) Data Science only — insight for humans, no learned model, arguably not "AI." (3) AI + ML + Deep Learning — speech-to-text is a deep-network task on unstructured audio. (4) Not ML and not really AI in the modern sense — hand-coded rules; at most trivially "symbolic," and it never changes with data.

Exercise 2 · The deliverable test

Take a project you've seen or imagined. Ask: is the deliverable a model that predicts, or an insight that informs a human decision? Explain how your answer places it as ML-leaning vs Data-Science-leaning — and note if it's genuinely both.

Exercise 3 · Justify the scope

In two or three sentences, explain to a peer why a course that stops at classical ML is still the right foundation for eventually understanding LLMs — using the words "features," "loss," and "interpretability."

One-paragraph summary to lock it in

AI, ML, Neural Networks, and Deep Learning form a strict nesting — each a smaller, more specific way of doing the one above it, with the defining moves being "learn from data," "use a network of neurons," and "make it deep." Data Science is not in that nesting; it overlaps ML but aims at insight for human decisions rather than a deployed model. Generative AI and LLMs sit two levels below where we stop, inside Deep Learning. The practical payoff is a decision rule — baseline with interpretable classical ML, reach for Deep Learning only when raw, abundant, unstructured data demands it — and that classical level is exactly, and deliberately, where this course lives.