Customize this training

Get a certificate — $16.67$5

Machine Learning Foundations: Core Concepts Without the Code

Machine Learning Foundations: Core Concepts Without the Code

Understand how machine learning models learn from data, how they are trained, tested and measured, and how to spot overfitting, data leakage, bias and drift before trusting a model.

What you will learn

  • Explain what machine learning is, how it differs from rule-based software and when it is the right tool.
  • Distinguish supervised, unsupervised and reinforcement learning, and classification from regression.
  • Prepare data correctly: features and labels, data quality, train/validation/test splits and leakage.
  • Describe how a model is trained and recognise underfitting, overfitting and the main model families.
  • Evaluate a model with the right metrics: confusion matrix, precision, recall, error measures and baselines.
  • Identify bias, drift and governance needs, and apply basic responsible-ML checks before and after deployment.

Modules

  1. What machine learning is and when to use it
  2. Types of machine learning
  3. Data: features, labels, quality and splits
  4. Training, generalisation and model families
  5. Measuring model performance
  6. Responsible ML: bias, drift and governance

Course facts

Level
Awareness
Estimated study time
about 45 minutes
Final assessment
24 questions · Pass mark 80%
Available in
English, Français, العربية
Price
Free. Optional verifiable certificate: $5
Last updated
Published by
AMAADOR ACADEMY

Who this course is for

Managers, analysts, project and product staff, students and any professional who works with data scientists or buys, uses or reviews ML-based tools. No vendor or programming language is assumed.

Prerequisites

None. Comfort with percentages and simple tables helps. No coding or advanced mathematics.

Course content

1. What machine learning is and when to use it

In classic software a programmer writes the rules: if the invoice is over 10,000 and the supplier is new, flag it. In machine learning (ML) nobody writes those rules by hand. You give the computer many examples, and an algorithm finds the patterns that link the inputs to the outcome. The result is a model: a mathematical function that takes new inputs and produces a prediction.

ISO/IEC 22989 places ML inside the wider field of artificial intelligence. The OECD defines an AI system as a machine-based system that infers, from the input it receives, how to generate outputs such as predictions, content, recommendations or decisions. Most AI products you meet today are built on ML.

TermWhat it means
Artificial intelligenceThe broad field: systems that produce predictions, content, recommendations or decisions
Machine learningThe part of AI where behaviour is learned from data instead of hand-written rules
Deep learningML with many-layered neural networks; powers image, speech and language models
Generative AIModels that produce new text, images or code, such as large language models

2. Types of machine learning

ML approaches are grouped by what the training data contains and how the model gets feedback.

Four families you will hear about in every ML discussion.

Supervised learning uses examples that come with the right answer, called the label. Past loan applications marked repaid or defaulted, photos tagged cat or dog. The model learns to predict the label for new cases. Most business ML is supervised.

Unsupervised learning has no labels. The algorithm looks for structure on its own: groups of similar customers (clustering), unusual transactions (anomaly detection) or a simpler summary of many variables (dimensionality reduction). A person must still interpret what the groups mean.

Reinforcement learning trains an agent that takes actions in an environment and receives rewards or penalties. Over many trials it learns a strategy that maximises reward. It is used for games, robotics and some control problems, and in fine-tuning language models with human feedback.

Self-supervised learning creates labels from the data itself, for example hiding the next word in a sentence and asking the model to predict it. This is how large language models are pre-trained, and it is the basis of most generative AI.

3. Data: features, labels, quality and splits

A dataset is a table. Each row is an example (a customer, a machine-hour, an email). Each input column is a feature. In supervised learning one column is the label or target: the answer you want to predict.

Tenure (months)Support callsPlanChurned? (label)
35BasicYes
410PremiumNo

Numbers such as tenure are numerical features. Values such as plan type are categorical features; they must be encoded as numbers before most models can use them, for example with one column per category (one-hot encoding). Large numeric ranges are often normalised so that one feature does not dominate just because its unit is bigger.

The saying garbage in, garbage out is literally true. A model will faithfully learn every error in its data. Check these points before any training:

  • Missing values — empty cells, or codes like 0 or 999 that really mean unknown.
  • Wrong or noisy labels — a defect recorded as OK teaches the model the wrong lesson.
  • Duplicates — the same record twice inflates its weight and can leak into the test set.
  • Outliers — a typing error of 10,000 kg instead of 10.0 kg.
  • Representativeness — does the data cover all the sites, seasons, customer groups and conditions the model will meet?
  • Class imbalance — if only 1% of cases are fraud, the model sees very few positive examples.

4. Training, generalisation and model families

Every model has parameters (also called weights): numbers it can adjust. A straight-line model that predicts delivery time from distance has two: a slope and a starting value. A large neural network can have billions.

Training is a loop that searches for the parameter values that make the fewest and smallest mistakes on the training data.

  1. Predict — the model makes predictions on a batch of training examples.
  2. Measure the loss — a loss function scores how wrong the predictions are. For regression a common choice is squared error; for classification, log loss (cross-entropy).
  3. Adjust — gradient descent works out in which direction each parameter should move to reduce the loss, and moves it a small step. The size of that step is the learning rate.
  4. Repeat — one full pass over the training data is an epoch. Training is often stopped once the loss stops improving on the validation set (early stopping).

Think of walking down a foggy hill: you cannot see the valley, but you can feel the slope under your feet and take a step downhill. A learning rate that is too large makes you overshoot and bounce around; one that is too small makes training very slow.

5. Measuring model performance

For classification, start with the confusion matrix. It counts the four possible outcomes when a model says positive (for example, fraud) or negative.

Actually positiveActually negative
Predicted positiveTrue positive (TP)False positive (FP), a false alarm
Predicted negativeFalse negative (FN), a missTrue negative (TN)

From it come the key metrics:

  • Accuracy = (TP + TN) / all cases. Share of correct predictions.
  • Precision = TP / (TP + FP). When the model raises an alert, how often is it right?
  • Recall (sensitivity) = TP / (TP + FN). Of all real positives, how many did it catch?
  • F1 score = the harmonic mean of precision and recall; useful when you need one number that balances both.
Typical case: 1,000 transactions, 10 of them fraud. A model that always answers not fraud scores 99% accuracy · it catches zero fraud, so recall is 0% · with imbalanced classes, report precision and recall, not accuracy alone.

Precision and recall pull against each other through the decision threshold. Lower the threshold and the model flags more cases: recall rises, but more false alarms cut precision. Raise it and the reverse happens.

6. Responsible ML: bias, drift and governance

A model learns from past data. If the past was unfair, incomplete or skewed, the model can repeat and even scale that pattern. This is not a rare edge case; it is one of the main risks that NIST’s AI Risk Management Framework asks organisations to manage.

  • Historical bias — the data faithfully records past unequal decisions.
  • Representation bias — some groups, regions or conditions are missing or rare in the data.
  • Measurement bias — the label is a poor proxy, for example arrests used for crime, or past spending used for need.
  • Proxy features — removing a sensitive attribute is not enough if postcode, school or first name stand in for it.
Case: Reuters reported in 2018 that Amazon dropped an experimental CV-screening model · trained on a decade of CVs mostly from men, it downgraded CVs containing the word women’s · Amazon said the tool was never used by its recruiters to evaluate candidates · the bias showed up when the team examined how the tool rated candidates; checking outcomes by gender before any use should be standard practice.

The full lesson continues in the interactive course on this page.

Sources and references

Conceptual introduction only. It does not make you a data scientist and does not replace the testing, validation, legal review or governance your organisation requires before deploying a model. Regulatory references (EU AI Act, NIST AI RMF) are summarised for awareness, not legal advice.

All courses · AI & Data - Free courses, certificate $5 | AMAADOR ACADEMY