Computer Science · Topic 10

Artificial Intelligence Fundamentals

Explore how artificial intelligence systems learn from data, recognize patterns, make predictions, generate content, and support decision-making.

1. What Is Artificial Intelligence?

Artificial intelligence (AI) is a field of computer science concerned with creating systems that perform tasks commonly associated with abilities such as perception, language processing, reasoning, planning, and decision-making.

Examples of AI

Speech recognition and voice assistants.
Image recognition and object detection.
Recommendations for music, videos, and products.
Navigation and route prediction.
Language translation and conversational systems.
Scientific data analysis and forecasting.

AI is a broad field. Not every automated program uses AI, and not every AI system learns from data after it has been deployed.

2. AI, Machine Learning, and Deep Learning

These terms are related, but they do not mean exactly the same thing.

Artificial Intelligence The broad field of building systems that perform tasks associated with intelligent behaviour.

Machine Learning A branch of AI in which systems learn patterns from data to make predictions, decisions, or representations.

Deep Learning A branch of machine learning that uses neural networks with multiple processing layers.

Artificial Intelligence
        |
        └── Machine Learning
                 |
                 └── Deep Learning

This diagram shows a common conceptual relationship. AI also includes approaches that do not rely on machine learning, such as rule-based reasoning and some search and planning methods.

3. How Does Machine Learning Work?

Machine learning uses data and an algorithm to build a model that can perform a task, such as classifying an image or estimating a value.

Collect data
     |
     v
Prepare and inspect data
     |
     v
Choose a model and learning method
     |
     v
Train the model
     |
     v
Evaluate on separate data
     |
     v
Deploy, monitor, and improve

Important terms

Dataset: A collection of examples used for analysis or learning.
Feature: An input property used by a model.
Label: The target answer provided for a training example in supervised learning.
Model: A learned or constructed representation used to produce outputs.
Training: The process of adjusting model parameters or structure according to a learning method.
Inference: Using a trained model to produce a prediction or output.

A model's output depends on its design, training data, and input. Training does not automatically guarantee that its predictions will be accurate.

4. Main Types of Machine Learning

Supervised learning

The model learns from examples containing inputs and known target answers.

Example: Training a model on labelled pictures of cats and dogs so it can classify new pictures.

Unsupervised learning

The model looks for structure or patterns in data without relying on labelled target answers for every example.

Example: Grouping customers into clusters based on similarities in their activity.

Reinforcement learning

An agent learns to choose actions by interacting with an environment and receiving rewards or penalties according to a defined objective.

Example: Training an agent to play a game by rewarding actions that improve its long-term score.

Type	Learning signal	Example task
Supervised	Known target answers	Classifying emails as spam or not spam
Unsupervised	Patterns in data	Grouping similar documents
Reinforcement	Rewards from actions	Learning a game-playing strategy

5. Neural Networks

An artificial neural network is a machine-learning model built from connected computational units commonly called neurons. These units combine inputs using adjustable parameters and activation functions.

Input layer
     |
     v
Hidden layer(s)
     |
     v
Output layer

Layers in a simple network

Input layer: Receives the model's input features.
Hidden layers: Transform information using learned parameters and activation functions.
Output layer: Produces the final prediction or other result.

Training a neural network

During training, the network's predictions are compared with a training objective. An optimization method adjusts its parameters to reduce the chosen loss function.

A common training method for neural networks uses backpropagation to calculate gradients and an optimizer to update parameters.

Neural networks are inspired loosely by biological neurons, but artificial neurons are mathematical computational units, not realistic copies of brain cells.

6. Generative AI

Generative AI refers to AI systems that produce new content, such as text, images, audio, video, or computer code.

Examples of generated content

Text Explanations, summaries, drafts, and dialogue.

Images Illustrations, concept art, and visual designs.

Audio Speech, music, and sound effects.

Code Code suggestions, examples, and explanations.

Large language models

A large language model (LLM) is trained on large amounts of text and other suitable data to model patterns in language. Many LLMs generate text by predicting likely next tokens based on the context and their learned parameters.

A token may represent a word, part of a word, punctuation, or another unit of text. Tokens are not always the same as words.

Fluent language is not proof of factual accuracy. Generative AI can produce incorrect claims, invented references, or misleading explanations, so important information should be checked.

7. Algorithms Used in AI

Different problems require different algorithms. No single method is best for every AI task.

Linear regression: Estimates numerical values using a linear relationship between inputs and output.
Decision trees: Make predictions through a sequence of feature-based decisions.
K-means clustering: Groups data into a chosen number of clusters according to distance-based criteria.
Nearest neighbours: Predicts or classifies using similar examples in a dataset.
Search algorithms: Explore possible states or paths to find a solution.
Neural networks: Learn complex patterns through layers of parameterized computations.

Choosing an algorithm involves considering the task, amount and quality of data, computational cost, interpretability, and how success will be measured.

8. Data Quality and Model Evaluation

The quality of data strongly influences what a model learns. Missing values, incorrect labels, unbalanced examples, and unrepresentative samples can lead to poor results.

Training, validation, and test data

Training set: Used to fit the model.
Validation set: Helps select models or tune settings.
Test set: Used for a final evaluation on data kept separate from training and model selection.

Overfitting and underfitting

Overfitting The model learns details or noise in the training data and performs poorly on new examples.

Underfitting The model fails to capture important patterns, performing poorly even on the training task.

Common evaluation measures

Accuracy: The proportion of predictions that are correct.
Precision: Among predicted positive cases, the proportion that are actually positive.
Recall: Among actual positive cases, the proportion correctly identified.
Mean squared error: The average of squared differences between predicted and actual numerical values.

Accuracy alone can be misleading when classes are highly imbalanced. The best evaluation measure depends on the real-world consequences of different kinds of errors.

9. AI Applications

Healthcare Supporting image analysis, research, and some clinical workflows under appropriate professional oversight.

Education Helping create practice material, explanations, and learning support.

Science Finding patterns in complex datasets and assisting research.

Agriculture Supporting crop monitoring, plant-disease detection, and resource planning.

Transport Estimating traffic, planning routes, and assisting driver-support systems.

Accessibility Supporting speech recognition, captions, translation, and image descriptions.

AI is most useful when its performance is appropriate for the task and its limitations are understood by the people using it.

10. Limitations and Responsible AI

AI systems can be powerful, but they also have limitations. Responsible development includes evaluating their effects on people and society.

Important challenges

Bias: Data or design choices can produce systematically unfair results.
Privacy: Personal data may be collected, inferred, or exposed inappropriately.
Reliability: A system can make errors, including confident-looking errors.
Explainability: It may be difficult to understand why a complex model produced a particular output.
Security: Systems may be manipulated or exposed to malicious inputs.
Resource use: Training and operating some models can require substantial computing power and energy.
Human oversight: High-impact decisions may require meaningful human review and accountability.

Good practices when using AI

Verify important claims using reliable sources.
Avoid entering confidential or sensitive information into services unless their data practices permit it.
Consider who could be unfairly affected by a model's errors.
Follow relevant rules for privacy, copyright, and academic integrity.
Use qualified human judgment for important medical, legal, financial, or safety decisions.

11. Knowledge Check

Try answering each question before revealing the answer.

1. What is the relationship between AI, machine learning, and deep learning?

AI is the broad field, machine learning is one branch of AI, and deep learning is a branch of machine learning that uses neural networks with multiple layers.

2. What is the difference between supervised and unsupervised learning?

Supervised learning uses examples with known target answers. Unsupervised learning looks for patterns or structure without requiring target labels for every example.

3. What does a model learn during training?

Depending on the method, it adjusts parameters or constructs a representation that captures useful patterns from the training data.

4. What is generative AI?

AI that generates content such as text, images, audio, video, or code.

5. What is overfitting?

When a model fits the training data too closely, including noise or accidental details, and struggles to generalize to new data.

6. Why should training and test data be kept separate?

A separate test set helps estimate how well the model performs on unseen data, provided it is not repeatedly used to tune the model.

7. Why should AI-generated information be verified?

AI systems can produce inaccurate or misleading outputs even when the responses sound convincing.

Key Takeaways

AI is a broad field focused on systems that perform tasks associated with intelligent behaviour.
Machine learning uses data to learn patterns; deep learning uses multilayer neural networks.
Supervised, unsupervised, and reinforcement learning use different learning approaches.
Generative AI can create text, images, audio, video, and code.
Data quality and careful evaluation affect model reliability.
Bias, privacy, security, and resource use are important AI challenges.
AI should be used thoughtfully, with verification and appropriate human oversight.