Unit 1: Foundations of Artificial Intelligence

I. Orientation: Intelligence as Goal-Directed Computation

Artificial Intelligence (AI) is the study and engineering of systems that perceive their environment, represent knowledge, reason, learn, and act to achieve goals. The modern field is conventionally associated with the 1956 Dartmouth workshop, although its foundations include logic, statistics, neuroscience, control theory, and computing.

• Goal-directed behaviour: An AI system selects actions that improve the achievement of a stated objective.
• Perception and action: Sensors provide information about the world; actuators or software outputs produce effects.
• Knowledge and uncertainty: Systems use data, rules, models, or probabilities because real environments are incomplete and noisy.
• Adaptation: Learning systems improve performance from examples or feedback rather than relying only on fixed instructions.
• Evaluation: AI quality is measured through task-specific criteria such as accuracy, reward, safety, latency, fairness, or cost.
• Human responsibility: Technical capability does not remove the need for human oversight, accountability, and ethical constraints.

II. Introduction to AI and its Evolution — From Rules to Learning Systems

AI developed from symbolic reasoning toward statistical learning and large-scale neural models. Each phase expanded the kinds of problems machines could handle.

A. Introduction to AI and its evolution

This topic explains AI’s historical development and the changing methods used to reproduce intelligent behaviour.

• Early foundations: Alan Turing’s 1950 paper asked whether machines could think and proposed an imitation game as a behavioural test.
• Symbolic AI: Early programs represented knowledge with logic and rules, such as IF fever AND cough THEN possible_infection.
• Expert systems: Systems such as MYCIN used hundreds of domain rules for medical advice, but were difficult to maintain outside narrow settings.
• AI winters: Limited computing power, insufficient data, and unrealistic promises contributed to reduced funding during periods such as the 1970s and late 1980s.
• Statistical learning: From the 1990s onward, algorithms learned patterns from datasets using methods such as decision trees, support vector machines, and probabilistic models.
• Deep learning era: GPUs, large datasets, and improved algorithms enabled neural networks to achieve major results in image recognition, speech processing, and translation.
• Current direction: Foundation models can perform multiple language, vision, coding, and generation tasks, but still require evaluation, governance, and domain supervision.

III. Types of Artificial Intelligence — Scope and Capability

AI can be classified by the range of tasks it performs or by its theoretical level of intelligence. These categories describe capability, not necessarily implementation.

A. Types of Artificial Intelligence

This classification distinguishes practical systems from hypothetical forms of machine intelligence.

• Narrow AI: A system designed for a specific task, such as spam detection, route planning, or face matching; nearly all deployed AI is narrow AI.
• General AI: A proposed system able to learn and transfer knowledge across diverse intellectual tasks at a human-like level; no verified general AI system currently exists.
• Superintelligence: A hypothetical intelligence exceeding human capability across virtually all domains; it remains a theoretical concept.
• Reactive systems: These respond to current inputs without storing meaningful past experience, as in a basic rule-based controller.
• Limited-memory systems: These use historical observations or trained parameters; autonomous driving and recommendation systems commonly fit this category.
• Learning capability: Classification should distinguish a system’s task range from its learning method: a narrow system can still use advanced deep learning.

IV. AI, ML, Deep Learning and Data Science — Related but Distinct Fields

These disciplines overlap but answer different questions. AI is the broad objective, while machine learning is one approach to achieving it.

A. AI vs ML vs Deep Learning vs Data Science

This comparison separates the scope, methods, and outputs of the four fields.

• Artificial Intelligence: The broad field of creating systems that perform tasks associated with intelligence, including planning, reasoning, perception, language, and decision-making.
• Machine Learning (ML): A method in which a model learns a relationship from data. In supervised learning, the model estimates a function such as:

  y_hat = f_theta(x)

Here, x is an input, y_hat is the prediction, and theta represents learned parameters.

• Deep Learning: ML using multilayer neural networks. A convolutional neural network may learn visual features from pixels, while a transformer processes relationships among tokens.
• Data Science: A broader practice combining statistics, programming, domain knowledge, data engineering, visualization, and experimentation to produce useful conclusions or decisions.
• Relationship: Deep learning is a subset of ML; ML is a major subset of AI; data science overlaps with all three but also includes non-AI analysis.
• Concrete distinction: A dashboard reporting monthly sales is data science but not necessarily AI; a rule-based chess engine is AI but not ML; a neural image classifier is both ML and deep learning.

V. Problem-Solving Using AI — Search, Knowledge and Decisions

AI problem-solving represents a task formally and searches for actions that transform an initial situation into a desired goal state.

A. Problem-solving using AI

This topic presents the standard formulation of an AI problem as states, actions, goals, and costs.

• Problem definition: Specify an initial state, available actions, transition model, goal test, and path-cost function.
• State-space search: A route problem can represent cities as states and roads as actions; a solution is a sequence such as A -> B -> D.
• Uninformed search: Breadth-first search explores shallow states first and finds a shortest path when every action has equal cost; depth-first search uses less memory but may miss the shortest solution.
• Informed search: A* evaluates:

  f(n) = g(n) + h(n)

g(n) is the cost from the start to node n; h(n) estimates the remaining cost. With an admissible heuristic that never overestimates, A* can find an optimal path.

• Knowledge representation: Facts, rules, graphs, ontologies, and embeddings allow a system to store relationships and use them during reasoning.
• Decision under uncertainty: Probabilities and expected utility help choose actions when outcomes are not guaranteed, such as selecting a diagnosis from symptoms.

B. AI Development Lifecycle

This lifecycle turns a problem idea into a deployed and monitored AI system.

• Problem framing: Define users, objective, constraints, and success metrics; “reduce fraud” must become measurable precision, recall, cost, or prevented loss.
• Data preparation: Collect, label, clean, represent, and split data into training, validation, and test sets while preventing leakage.
• Model development: Select an algorithm, train parameters, tune hyperparameters, and compare against a simple baseline.
• Evaluation: Test accuracy alongside robustness, bias, explainability, latency, security, and performance on relevant subgroups.
• Deployment: Integrate the model into an application, establish access controls, logging, rollback, and human escalation.
• Monitoring and maintenance: Track data drift, concept drift, errors, and changing user behaviour; retraining is required when performance degrades.
• Governance: Document intended use, limitations, datasets, model versions, approvals, and incident responses throughout the lifecycle.

VI. Intelligent Agents and Agent Environment — Perception to Action

An intelligent agent is an entity that perceives through sensors and acts through actuators to maximize performance relative to a specified measure.

A. Intelligent Agents and Agent Environment

This topic defines the agent-environment relationship and the properties that determine suitable agent designs.

• Agent function: It maps a percept history to an action: f: P* -> A, where P* is the sequence of percepts and A is the set of possible actions.
• Performance measure: A vacuum agent may be scored by cleanliness, energy use, time, and collision avoidance rather than movement alone.
• PEAS description: Performance measure, environment, actuators, and sensors specify an agent’s task. A taxi agent uses safety and travel time, roads and traffic, steering and brakes, and cameras or GPS.
• Simple reflex agent: Selects an action from the current percept using condition-action rules; it is fast but fails when the current percept hides relevant history.
• Model-based agent: Maintains an internal state describing unobserved aspects of the environment.
• Goal-based and utility-based agents: Goal-based agents seek a desired state; utility-based agents rank alternatives, for example balancing speed against fuel consumption.
• Learning agent: Improves through a performance element, learning element, critic, and problem generator.
• Environment properties: Environments may be fully or partially observable, deterministic or stochastic, episodic or sequential, static or dynamic, and discrete or continuous.

VII. Applications of AI in Robotics, Healthcare, Manufacturing and Smart Cities — AI in the Physical World

AI applications combine perception, prediction, planning, and control with domain-specific data and safety requirements.

A. Applications of AI in Robotics, Healthcare, Manufacturing and Smart Cities

This topic shows how AI supports decisions and automation across major sectors.

• Robotics: Computer vision identifies objects, path planning avoids obstacles, and reinforcement learning can optimize actions; a warehouse robot may use LiDAR, cameras, localization, and a motion controller.
• Healthcare: AI assists image analysis, triage, drug discovery, and personalized risk prediction. A model detecting pneumonia from X-rays must be validated across hospitals and should support, not silently replace, clinical judgment.
• Manufacturing: Predictive maintenance estimates failure risk from vibration, temperature, and current measurements. Detecting abnormal vibration before a bearing fails can reduce unplanned downtime.
• Smart cities: AI can optimize traffic signals, public transport, energy distribution, waste collection, and emergency response using sensor and infrastructure data.
• Operational constraints: Robotics requires real-time safety; healthcare requires privacy and clinical validation; manufacturing requires reliability; smart cities require public accountability and protection against surveillance abuse.
• Human role: High-impact deployments need override mechanisms, audit logs, fallback procedures, and clear responsibility for errors.

VIII. Ethics in AI and Responsible AI — Trustworthy Deployment

AI ethics examines how systems affect people, institutions, and the environment. Responsible AI converts ethical principles into requirements, controls, and measurable practices.

A. Ethics in AI and Responsible AI

This topic focuses on preventing harm while preserving useful innovation.

• Fairness: A model should not produce unjustified differences across groups. Comparing false-positive rates for different demographic groups can reveal disparate impact.
• Transparency and explainability: Users should know when AI is involved, what data informs a decision, and which factors influenced an output where explanation is technically and legally appropriate.
• Privacy: Data minimization, consent, access controls, anonymization, and encryption reduce risks from collecting sensitive health or location data.
• Accountability: Named owners, documentation, audit trails, impact assessments, and appeal processes ensure that responsibility remains with organizations and people.
• Safety and robustness: Models should be tested against distribution shifts, adversarial inputs, failures, and unsafe edge cases before and after deployment.
• Human oversight: Human-in-the-loop review is important for high-impact decisions, but reviewers need authority, time, training, and meaningful information.
• Sustainability: Training and operating large models consume electricity and hardware resources; efficiency is an ethical and economic consideration.
• Responsible practice: Use representative data, document limitations, monitor outcomes, report incidents, and stop or redesign systems whose harms cannot be controlled.

IX. Introduction to Generative AI — Models That Create Content

Generative AI produces new text, images, audio, video, or code by learning patterns in large datasets and sampling outputs from a learned distribution.

A. Introduction to Generative AI

This topic introduces the operating principle, common architectures, uses, and limitations of generative systems.

• Generative objective: A language model estimates the next-token probability:

  P(x_t | x_1, x_2, ..., x_(t-1))

x_t is the next token, and the preceding tokens form the context.

• Training: During pretraining, model parameters are adjusted to reduce prediction error over large datasets; later fine-tuning or preference optimization can shape behaviour.
• Transformers: Attention allows each token to assign different importance to other tokens, supporting long-range relationships in language and multimodal inputs.
• Other architectures: Generative adversarial networks use a generator and discriminator; diffusion models gradually denoise random noise to create images or other media.
• Prompting and grounding: A prompt specifies an instruction and context; retrieval-augmented generation supplies external documents to reduce reliance on internal model memory.
• Applications: Systems can draft reports, summarize text, generate software, create designs, translate languages, and provide conversational interfaces.
• Limitations: Outputs may contain hallucinations, bias, copyright concerns, privacy leaks, prompt-injection vulnerabilities, or confidently incorrect code.
• Responsible use: Verify factual and high-impact outputs, protect confidential inputs, disclose generated content where relevant, and retain human ownership of final decisions.