AI, Machine Learning and Deep Learning

Objective: locate deep learning, understand what it changes and avoid the "neural network for everything" reflex. Duration: 35 minutes.

1. The nested dolls

Artificial intelligence covers the techniques that let a machine perform a task associated with intelligence. Machine learning is a part of AI in which rules are adjusted from data. Deep learning is a part of machine learning that learns with networks made of several stages of transformation.

text

AI ⊃ Machine Learning ⊃ Deep Learning

The word deep describes the number of representation steps, not depth of thought. An image can successively become: pixels → edges → textures → parts → object.

2. What deep learning brings

Classical methods often require a human to build the useful variables. A deep network can learn part of that representation directly from raw data. This property is particularly useful for images, audio, text and complex signals.

Situation	First reasonable candidate
Small table of 3,000 rows	tree, boosting, regression
Millions of labeled images	CNN or Vision Transformer
Text with a pretrained model	fine-tuned Transformer
Exact and stable business rule	classical programming

3. The price to pay

Deep networks often demand more data, more compute and more experimentation. They are sensitive to the protocol, initialization and learning rate. A larger model is neither automatically more accurate nor automatically more useful.

Common mistakes

Confusing demo performance with business value.
Comparing an optimized network to a neglected classical baseline.
Using the test set while making architecture choices.
Believing that "deep" means "autonomous" or "conscious".

Quick check

Is deep learning a family distinct from machine learning?
Why is it strong on unstructured data?
Give a case where it is likely overkill.

Answers

No, it is a subset of ML. 2. It learns hierarchical representations from raw data. 3. For example a small table or an exact rule.

Mastery activity — Decision memo

Pick three cases: a churn table, chest X-rays and an exact tax rule. For each, propose the simplest solution, a baseline, the expected proof and a stopping criterion. Deliverable: a one-page table. Success: no solution is chosen only because it is "modern".

Decision map: which paradigm to choose?

Reading: deep learning is not an automatic choice. The nature of the data, the amount of supervision, the compute budget and the operating cost determine the solution. The main trap is picking the most complex architecture before establishing a measurable baseline.