Artificial Intelligence - Introduction and Basics: formula sheet

Key formulas

Working definition of AI

AI = machine performing tasks that normally need human intelligence (learning + reasoning + perception + decision-making)

Write this in your own words in the first line of any definition answer.

Traditional software logic

Input + Rules (written by programmer) → Output

Behaviour is fixed and explicit; same input gives same output.

AI / machine learning logic

Input data + Expected outputs → Learned rules (model) → Predictions on new data

Behaviour comes from training data and may change when retrained.

Turing Test

Machine passes if a human judge cannot reliably distinguish its responses from a human's

Proposed by Alan Turing in 1950; a test of behaviour, not of true understanding.

Classification by capability

Narrow AI → General AI → Super AI

Narrow: one task. General: human-level across tasks. Super: beyond human. Only narrow AI exists today.

Classification by functionality

Reactive machines → Limited memory → Theory of mind → Self-aware

Capability rises along the list. The first two exist; the last two are theoretical.

Rough mapping between the two

Reactive and limited memory ≈ Narrow AI; Theory of mind ≈ General AI; Self-aware ≈ Super AI

Use as a guide and say 'broadly'. Sources vary on exact alignment.

Memory test

No memory = reactive; uses past data = limited memory

Quick way to place a given example.

Nesting of terms

AI ⊃ Machine Learning ⊃ Deep Learning

Deep learning is a subset of ML, and ML is a subset of AI. Not every AI system uses ML.

Supervised learning

Labelled data (input + known output) → model learns mapping → predicts output for new input

Output is a category (classification) or a number (regression).

Unsupervised learning

Unlabelled data → model finds hidden patterns, groups or outliers

Main tasks: clustering, anomaly detection, association.

Reinforcement learning

Agent takes action → environment gives reward or penalty → agent improves its policy

Learning is by trial and error, aiming at maximum long-term reward.

Neural network structure

Input layer → hidden layer(s) → output layer

Deep learning means many hidden layers.

Neuron computation

Output = activation( Σ (weight × input) + bias )

Training adjusts weights and bias to reduce error.

NLP

Text or speech → language understanding → classification, extraction, summary or reply

Input is language. Use for chatbots, contract review and sentiment analysis.

Computer vision

Image or video → pattern recognition → identification, reading or detection

Input is visual. Use for OCR, face match and defect detection.

Expert system

Knowledge base (rules) + inference engine + facts → advice

Rule-based and explainable. It does not learn from data by itself.

Generative AI

Learned patterns from training data + prompt → new content

Creates content. Output needs human verification because of hallucination.

Robotics vs RPA

Robotics = physical machine; RPA = software bot following rules

A common trap. RPA is not a physical robot.

AI lifecycle sequence

Problem definition → Data collection → Data preparation → Algorithm selection and training → Testing and validation → Deployment → Monitoring and maintenance

Stage names vary between textbooks. Keep the order and the purpose of each stage. Monitoring often feeds back into retraining.

Data split

Dataset = Training set + Validation set + Test set

Training set teaches the model, validation set tunes it, test set gives a final unbiased check. Some projects use only training and test sets.

Core relationship

Model = Algorithm applied to Training data

The algorithm is the method. The model is what it learns. Same algorithm with different data gives a different model.

Data quality principle

Quality of output depends on quality of input data (garbage in, garbage out)

Quality means accuracy, completeness, consistency, timeliness and representativeness.

Core benefits of AI

Automation + Accuracy + Speed + Scale + Insight + Availability

Use as a checklist for the advantages part of an answer.

Core risks of AI

Bias + Opacity + Privacy + Security + Accountability + Job displacement

Cover all six in a full answer, with one example each.

Source of algorithmic bias

Biased or incomplete data (or design) → biased model output

Bias enters through data, labelling choices or model design, not only through the algorithm.

Black box problem

Complex model → output without an explainable reason

The remedy is explainability, documentation and human review.

Risk response

Identify → Assess → Control → Monitor

Use this to structure the mitigation part of any answer.

Quick revision

• Artificial Intelligence means machines performing tasks that normally need human intelligence, such as learning, reasoning and decision-making.
• AI can be classified by capability (narrow, general, super) and by functionality (such as reactive and limited memory).
• Most AI in use today is narrow AI, built for a specific task.
• Machine learning lets systems learn patterns from data instead of following only fixed rules.
• Main machine learning styles are supervised, unsupervised and reinforcement learning.
• Deep learning is a subset of machine learning that uses multi-layer neural networks and usually needs large data and computing power.
• Key technologies include natural language processing, computer vision, speech recognition and robotics.
• The AI lifecycle runs from data collection and preparation to model training, testing, deployment and monitoring.
• The quality of an AI output depends heavily on the quality of the data and the design of the algorithm.
• Benefits include efficiency, accuracy and better decisions; risks include bias, lack of transparency, privacy harm and misuse.
• In answers, link each risk to a governance or compliance response.

Common mistakes

• Defining AI only as robots or chatbots Fix: Define AI as a field of systems that perform tasks needing human intelligence; robots and chatbots are only applications.
• Saying AI and traditional software are the same because both are programs Fix: State that traditional software follows fixed programmed rules while AI learns patterns from data and improves with training.
• Saying general AI or super AI already exists. Fix: State clearly that all deployed AI is narrow. General and super AI are hypothetical.
• Mixing up the two classifications and listing 'reactive machines' as a type of general AI. Fix: Write them separately: capability (3 types) and functionality (4 types). Then give the rough mapping.
• Treating AI, machine learning and deep learning as the same thing. Fix: Always state the nesting: deep learning is a subset of ML, which is a subset of AI. Some AI uses fixed rules and no learning.
• Saying unsupervised learning has no data or no training. Fix: Say it uses data without labels and finds structure itself.
• Treating RPA as robotics or as machine learning. Fix: Say RPA is rule-based software that automates repetitive digital tasks. Robotics involves a physical machine.
• Saying expert systems learn from data. Fix: Expert systems apply rules written by human experts. Learning from data belongs to machine learning.
• Using algorithm and model as the same thing. Fix: Write the definition: the algorithm is the learning method; the model is the trained result.
• Stopping the lifecycle at deployment. Fix: Always add monitoring, maintenance and retraining. Mention drift as the reason.

Exam tips

• Always begin with a clear definition; examiners award marks for it even in long answers.
• Learn a short timeline with years: 1950, 1956, expert systems, AI winters, machine learning, deep learning, generative AI.
• For distinguish questions, give at least five matched points and one example.
• Link AI basics to governance, data protection and cyber security in your conclusion, as this is a laws and practice paper.
• Write only the dates and milestones you are sure of; follow the ICSI study material for finer details.
• Draw a small two-column layout in words: capability types and functionality types. Examiners often reward a clear structure.
• Always say which types exist today and which are theoretical. This is a common mark-earning point.
• In case-based questions, classify the system in the facts first, then discuss the legal or risk issue that follows.