Chapter 1 — Introduction to Machine Learning

Learning from data

Bryan G. Pantoja-Rosero

What to expect?

overview over ML/AI
basic understanding of most important ML methods and fundamental concepts
experience how ML is done on a practical problem

What not to expect?

instantly becoming an expert AI specialist / data scientist / ML researcher / statistician / ML software expert / domain expert
learn about that newest hack you saw on blog xyz.

Course website

What is machine learning?

Machine learning (ML): algorithms that learn useful relationships from data.

Conventional programs: rules specified explicitly.
ML systems: some decision rules are estimated from examples.
Goal: useful predictions on previously unseen inputs.
Learning patterns ≠ human understanding or guaranteed correctness.
What information is available, what do we want to predict, and how can examples help us make that prediction?

Learning rules from data

Example: post-earthquake damage assessment.
Input: photographs of buildings.
Desired output: damaged or undamaged.
Hand-crafted approach: color, shape, texture, and threshold rules.
ML approach: learn parameters from labeled examples.

Question: Which patterns remain useful on new photographs?

Traditional programming versus learning

Rules versus learning

Training, testing, validation

Training: estimate model parameters

Given supervised examples,

\[ \mathcal D=\{(\mathbf x_i,y_i)\}_{i=1}^{n} \]

\(\mathbf x_i\): input; \(y_i\): known target/label.
Learning algorithm updates parameters \(\mathbf w\).
Trained model: \(f(\mathbf x;\mathbf w)\).

Training asks: How do we estimate a model from examples?

Training, testing, validation

Inference and evaluation are not training

Inference — predict on a new input

\[\hat y=f(\mathbf x_{\mathrm{new}};\mathbf w)\]

The correct target does not need to be available.

Evaluation — measure predictive performance

Compare predictions with known targets.
Prefer held-out, previously unseen data.

Three distinct questions: learn? predict? reliable?

Learning functions from input–output pairs

The most useful way to recognize a machine-learning problem is often to identify two related kinds of information: an input \(\mathbf x\) and a desired output \(y\).
A collection of matched examples \((\mathbf x,y)\) provides the evidence from which we can learn a mapping.

A first classification model

Post-earthquake damage assessment

Binary task: damaged vs. undamaged buildings

Input: building images.
Data: labeled photographs that belong to one of two classes.
Training: estimate a classification rule.
Inference: apply that rule to a new image.
Output: building class.
Underlying steps: pixels → numbers → score → class.

Building damage classification

What an update can—and cannot—do

A wrong prediction provides a signal to change weights.
Updating one example may affect predictions for others.
A perfectly separating hyperplane may not exist.
Related to stochastic subgradient learning for a perceptron loss.

Central idea: prediction error drives parameter adaptation.

Machine learning is interdisciplinary

Artificial versus biological neural networks

The name neural network is inspired by neuroscience.
Artificial neuron: numerical weighted operations.
Biological neuron: complex electrical and chemical processes.
Both concern adaptive systems—but mechanisms are not equivalent.

Why machine learning matters

A good model is not enough without reliable data and useful decisions.

Origins: learning from observations

John Snow and the 1854 London cholera outbreak

Pre-ML lesson: data patterns can guide scientific hypotheses.

Foundations of modern AI and ML

Durable ideas: numerical representations, learning rules, prediction.

What enabled recent progress?

Three interacting developments:

Computing capacity — vastly increased. Digital datasets — larger and more accessible. Algorithms — more effective and efficient.

Challenges of ML

Hype versus real-world reliability

AI has experienced cycles of enthusiasm and disappointment.
A convincing demonstration is not proof of reliable deployment.
Ask: What assumptions were made? On which data was it tested?
Performance must be judged in its application context.

Data, privacy, and fairness

Datasets reflect collection, labeling, and historical decisions.
Observed patterns can be biased or unrepresentative.
A model can reproduce unfair outcomes without an explicit unfair rule.
Privacy and data governance are part of system design.

Challenges of ML

Example: biased post-earthquake inspections

Evaluate coverage, representativeness, and consequences.

Challenges of ML

Interpretability and societal impact

Simple rules may be transparent; deep models can be harder to explain.
High-stakes domains need reliability and physical plausibility.
AI affects work, education, information, and access to services.
Potential misuse: misinformation and harmful automation.
Longer-term safety and governance questions remain under debate.

Building trustworthy ML

Follow the scientific method and evaluate reproducibly.
Document training data, assumptions, and limitations.
Use transparent benchmarks and independent validation.
Openness helps reproducibility—but balance with privacy, licensing, and safety.

Data modalities and modern applications

Different data can be represented numerically.
Task depends on input data and desired output.
Outcomes include classes, regions, numbers, sequences, and actions.

Next: different forms of data—and the problems they make possible.

Digital text makes large language models possible

Train on extensive collections of digital text.
Learn relationships between tokens and context.
Adapt for tasks: Q&A, translation, summaries, coding.
Powerful outputs can still be plausible but inaccurate.

Language models and generative AI

Large language models (LLMs)

Learn statistical patterns of human language.
Trained on large text collections with many parameters.
Many use next-token prediction as a central training objective.

Language-model capabilities over time

Source: Artificial Analysis, Frontier Language Model Intelligence, Over Time (accessed 12 October 2026). https://artificialanalysis.ai/#frontier-language-model-intelligence-over-time

Generative AI goes beyond text

Image generation: new images from text prompts.
Audio: synthetic speech and music.
Video: generated moving scenes.
Other representations: 3D data and multimodal content. Source: Max Roser (2022), The brief history of artificial intelligence: the world has changed fast — what might be next?, Our World in Data. https://ourworldindata.org/brief-history-of-ai

Generating plausible content ≠ being correct

Generative models learn patterns and produce new content.
Outputs may appear convincing but be factually wrong.
Validate outputs for scientific, technical, and safety-critical uses.
Responsible use requires knowing strengths and limitations.

From opportunities to meaningful ML problems

Ask before choosing an algorithm:

What decision or prediction would be useful?
Which inputs exist at prediction time?
Can we obtain reliable targets?
How will success be measured on unseen cases?
What are the costs of errors?

Example: infrastructure inspection and maintenance

Possible outputs: damage class, crack width, deterioration, priority.
Inputs: photographs, sensor data, weather, inspection history.
Labels: observed damage categories or measured conditions.
Evaluation: unseen structures, representation, consequences.

Engineering value depends on more than predictive accuracy.

From an application to a mathematical model

Inputs \(\mathbf x\) and outputs \(y\) define a learning problem.
A model \(f(\mathbf x;\mathbf w)\) makes predictions.
A loss function quantifies prediction errors.
Optimization adjusts parameters to reduce loss.
Generalization asks whether the model works on new cases.

Now it is your turn

Machine Learning Mini-Project

Open topic: any domain, dataset, input type, or suitable ML model.

Define a problem: choose or create a dataset; identify inputs and desired outputs.
Build a model: prepare data, train, and evaluate on unseen examples.
Demonstrate: provide a runnable prediction/test workflow (script or notebook is fine).

Submit

GitHub repository with a README explaining setup, data access, retraining, testing, and results.
2-page LaTex report: problem, approach, main findings, lessons learned, and relevance to your field.

Priority: clear problem + reproducible workflow + meaningful evaluation. No web app or novel algorithm required.

Coming next — regression and linear regression

Represent input–output relationships with mathematical functions.
Define predictions and measure errors with loss functions.
Use optimization to adjust model parameters.
Examine model fitting and generalization.

Simple regression models provide foundations for more advanced ML.

Acknowledgements and AI Disclosure

Acknowledgements

Developed with reference to CS-433 Machine Learning, École Polytechnique Fédérale de Lausanne (EPFL).
Adapted and extended for the present course, including civil engineering applications.
Independent teaching materials; no EPFL endorsement is implied.
Original EPFL course

Use of Generative AI

AI-assisted drafting, editing, examples, figures, and interactive visualizations.
The instructor is responsible for the final content and technical accuracy.
External materials are acknowledged separately.