7. Emerging technologies

Artificial intelligence is a branch of computer science focused on the simulation of intelligent behaviours by computers — behaviours we would call intelligent if a person did them, such as recognising a face, understanding a sentence, making a decision or learning from mistakes.

• The computer does not think like a human; it simulates intelligent behaviour by processing data.
• Most modern AI is machine learning: instead of a programmer writing every rule, the system works out its own rules from examples.

• Supervised learning — trained on labelled data: every example comes with the correct answer (photos labelled "cat" / "dog", emails labelled "spam" / "not spam"). The system learns to predict the label for new, unseen data.
• Unsupervised learning — given unlabelled data and finds patterns or groups by itself, e.g. grouping shoppers with similar buying habits so a shop can target them.
• Reinforcement learning — learns by trial and error: it takes actions, receives a reward for good outcomes and a penalty for bad ones, and adjusts to earn more reward (game-playing AI, a robot learning to walk).

• Recommendation systems — suggest videos, songs or products from your history and from people similar to you.
• Natural language processing (NLP) and large language models (LLMs) — understanding and producing human language: chatbots, translation, voice assistants. An LLM is trained on huge amounts of text and predicts the most likely next word, which lets it answer questions and write text.
• Generative AI — creates new content (text, images and video, music) from a prompt, based on patterns learned from training data.
• Computer vision — recognising objects, faces, text or road signs in images and video; links to robotics and automated systems, e.g. a self-driving car seeing pedestrians, a factory robot checking products, a fruit-picking robot judging ripeness.
• Predictive analytics — using past data to predict what is likely to happen: weather, stock demand, which customers may leave, risk of disease.

A neural network is a model loosely based on the brain, made of connected nodes (artificial neurons) arranged in layers:

1. Input layer — receives the data, one node per input value (e.g. the brightness of each pixel of an image).
2. Hidden layer(s) — one or more layers between input and output; each node combines the values it receives using weights and passes a result on. This is where patterns are found. More hidden layers = "deep learning".
3. Output layer — gives the result, e.g. one node per possible answer ("cat" 0.92, "dog" 0.08).

Every connection has a weight that says how important that input is. Learning means adjusting the weights.

The purpose of machine learning is to let a computer improve its performance at a task from data/experience without being explicitly programmed with every rule — useful when the rules are too complex to write by hand (recognising handwriting, spotting fraud).

How a neural network is used:

1. Training — training data is fed into the input layer; the network produces an output.
2. The output is compared with the correct answer and the error is measured.
3. The weights are adjusted to reduce the error; this repeats over thousands of examples.
4. Testing — the network is checked on data it has never seen.
5. Use — new data goes in, and the trained network produces a prediction or classification.

Exam questions describe a scenario and ask how AI could be used. Match the need to an application:

Always say what data the AI uses and what it does with it.

An AI system is only as fair as the data it was trained on and the people who designed it.

• Biased data — if the training data under-represents a group (e.g. few images of darker skin tones), the system works less accurately for that group.
• Historic bias — data that reflects past unfair decisions (e.g. past hiring favouring men) teaches the AI to repeat them.
• Design bias — the choice of which data to collect and what counts as "success" can favour some people.
• Effects: unfair loan, job or policing decisions; facial recognition mistakes; stereotypes in generated images.

Reducing bias: use large, varied, representative data; test results for different groups; keep humans checking important decisions; make the system's decisions explainable.

• Privacy — AI needs lots of personal data; how is it collected, stored and shared? Was consent given?
• Accountability — who is responsible when AI gets it wrong (a self-driving car crash, a wrong diagnosis): the developer, the owner or the user?
• Transparency — many systems are a "black box"; people affected may not know why a decision was made.
• Misinformation — generative AI can create convincing fake images, video ("deepfakes") and text.
• Copyright and ownership — models learn from other people's work; who owns the output?
• Over-reliance — people may trust AI answers without checking them; AI can state wrong facts confidently.
• Access — benefits may go mainly to those who can afford the technology (digital divide).

"Evaluate" questions need both sides and a conclusion.

• 3 classical bits hold one of 8 values at a time; 3 qubits can represent all 8 combinations at once. Adding qubits doubles the combinations each time.
• This lets a quantum computer explore a huge number of possibilities together, so some kinds of problem (searching huge numbers of combinations, simulating molecules, optimisation, breaking some encryption) could be solved much faster.
• It is not simply a faster PC: for everyday tasks (word processing, web browsing) a classical computer is better.

Advantages

• Can solve certain complex problems far faster than classical computers (e.g. modelling molecules for new medicines and materials, optimisation of routes and schedules, weather and climate modelling).
• Can process many possible combinations at the same time.
• Could create very secure communication (quantum encryption).

Disadvantages

• Very expensive to build and run.
• Qubits are fragile — they lose their state easily (heat, vibration, interference), causing errors.
• Often need extreme cooling to near absolute zero, using a lot of energy and space.
• Only useful for specific types of problem; not suitable for everyday use.
• Few programmers have the skills; few programs exist.
• Could break current encryption methods, threatening data security.