AI vs Machine Learning vs Deep Learning Explained

What you will learn

Draw the relationship between AI, ML, and deep learning.
Explain what is learned in a model.
Identify a neural network as a model family.
Choose a sensible starting method for a small task.
Separate a model's architecture from its task.

Before you begin

Read AI01 first so input, output, training, and inference are familiar.

What Is Artificial Intelligence? A Beginner’s Introduction

Three names at different scales

A shop wants to sort photographs of its products. One developer proposes AI, another says machine learning, and a third recommends deep learning. They may be describing the same project at different levels. AI is the broad field. Machine learning is a group of AI methods that improve model behavior through data. Deep learning is machine learning using neural networks with multiple learned layers.

The nesting is useful, but it is not a ranking from bad to good. AI can include search, planning, or explicit knowledge rules without using deep learning. A learned decision tree is machine learning without being a deep neural network. A deep image classifier is all three: AI, ML, and deep learning.

Also distinguish a model from a product. A product may contain a deep classifier, a database query, a login form, and a manually written price calculation. The AI label does not make every component a learning system.

Artificial intelligence
└── Machine learning
    └── Deep learning

What changes when software learns?

Start with a rule: place files larger than five megabytes into one folder. A person specifies the condition, and the computer applies it. If the threshold is wrong, the programmer edits it. This is ordinary programming, whether the surrounding application markets itself as AI or not.

For a learned product classifier, a person supplies examples such as images labeled shoe or bag. Training adjusts numerical parameters so predictions better match those labels. Humans still choose the task, examples, model family, and evaluation method. Learning does not remove design decisions; it changes where some behavior comes from.

A model learns statistical relationships, including unwanted shortcuts. If every shoe image has a white background and every bag image has a blue one, the model may exploit the background. Testing shoes against blue backgrounds reveals that problem. Behavior must be assessed, not assumed from the method.

Why layers matter

A neural network transforms numbers through connected layers. Each layer calculates new values from previous values using adjustable weights, biases, and usually nonlinear functions. Stacking these transformations can represent complicated relationships between inputs and outputs. The name neuron is an analogy; this is numerical computation, not a tiny biological brain.

For images, a network may learn useful visual representations at different stages. Early representations can respond to local patterns, while later representations combine information in more complex ways. This is an intuition, not a promise that every layer corresponds neatly to a human idea such as wheel or handle.

Deep learning can work well with images, sound, and language, but usually adds computational and data demands. A small table of measurements may be handled well by a simpler model. Nonlinear transformations allow networks to represent patterns that a single linear model cannot. Neural-network fundamentals.

Choose the smallest useful starting point

Suppose your actual task is separating photos by the date they were taken. Read the date stored in the file and use ordinary code. Training an image model would solve a harder problem than the one you have. If the task is recognizing product type from pixels, an existing trained image model may be reasonable, provided its categories and use conditions fit.

For a classroom project predicting plant height from recorded days and watering conditions, start with basic regression and clear evaluation. A larger network does not create missing information. It can fit accidental patterns more closely while providing worse predictions on new plants.

Compare candidates using the same held-out examples, measurement, and operating constraints. Include memory, speed, maintenance, and the cost of mistakes. The best method meets the need with acceptable evidence; it is not necessarily the method with the longest name.

Approach	Where behavior comes from	Example
Explicit rule	Conditions written by a person	Sort by file date
Machine learning	Parameters fitted from examples	Predict plant height
Deep learning	Many learned network layers	Recognize image content

Important terms

Machine learning: Methods that fit model behavior from data.
Deep learning: Machine learning using neural networks with multiple learned layers.
Parameter: A model value adjusted during fitting.
Architecture: The structural design of a model.
Baseline: A straightforward comparison method.
Held-out data: Examples reserved from fitting to assess performance.

Mini project: Sort five tasks by method

Consider calculating a bill, detecting a damaged leaf, grouping sounds, sorting by date, and predicting temperature.
For each, propose an explicit rule or learned approach and explain why.
For one learned task, specify the input, target output, and baseline.
Finish with a case where deep learning could help and one where it would add unnecessary complexity.

Common mistakes and debugging

Using AI and deep learning as exact synonyms: remember the nested relationship.
Assuming more layers guarantee better results: compare on unseen examples.
Calling all adjustable settings learned parameters: developers choose some settings before training.
Judging a whole app from one model: inspect surrounding rules and processing.

Independent challenge

Design a photo organizer combining file-date rules with a learned classifier. Explain what happens when the date is missing or the model is uncertain.

Check your understanding: 10 questions

Is every deep-learning model also machine learning?
Is every AI system a deep network?
What is a parameter?
Why might background color mislead a classifier?
How should two methods be compared?
In your own words, what does “Machine learning” mean?
In your own words, what does “Deep learning” mean?
In your own words, what does “Parameter” mean?
In your own words, what does “Architecture” mean?
In your own words, what does “Baseline” mean?

Quiz answers

Reveal all 10 answers after your attempt

Yes. Deep learning is a subset of machine learning.
No. AI includes other learned methods and approaches such as search and explicit reasoning rules.
A numerical model value adjusted during training.
It can correlate with labels without representing the product itself.
Use appropriate shared evaluation data and metrics, then consider operating constraints.
Methods that fit model behavior from data.
Machine learning using neural networks with multiple learned layers.
A model value adjusted during fitting.
The structural design of a model.
A straightforward comparison method.

Summary

AI is the broad field; ML learns from data; deep learning uses layered neural networks. Choose a method by the problem and evaluation evidence.

Continue learning

AI03 explains generative AI, a category defined by outputs rather than a single architecture.

Choose a connected learning path

Sources and further reading

Google: Neural networks

Prepared 2026-09-18. Draft — conceptual relationships source checked