Unit 1: Introduction to Artificial Intelligence

I. Orientation — The Study of Intelligent Machines

Artificial intelligence (AI) is the field of computing concerned with creating systems that perform tasks normally associated with human intelligence. These tasks include perceiving the environment, learning from information, reasoning about alternatives, understanding language, and taking appropriate action. The modern field was formally named at the Dartmouth workshop in 1956, although its intellectual roots extend through logic, mathematics, neuroscience, and early computing.

• Governing principle: An AI system receives inputs, processes them using representations and algorithms, and produces predictions, decisions, or actions.
• Core objective: The system should achieve a specified goal effectively, not merely imitate human appearance.
• Intelligence as adaptation: Intelligent behavior involves responding appropriately when conditions, information, or goals change.
• Role of data and knowledge: AI may use explicit rules, stored knowledge, statistical patterns, learned representations, or combinations of these.
• Evaluation convention: AI performance is assessed through task-based measures such as accuracy, planning success, speed, safety, robustness, and human usefulness.
• Important distinction: Automation follows fixed instructions, whereas AI generally handles uncertainty, variation, or complex patterns.

II. Concept of Intelligence — Meaning and Components

A. Definition and dimensions

Intelligence is the capacity of an agent to acquire information, learn from experience, reason about situations, solve problems, and act successfully toward goals. In AI, intelligence is usually treated as a functional capability rather than as a claim that a machine possesses human consciousness.

• Perception: An intelligent agent converts sensory data into useful information; for example, a vision system may identify a pedestrian in camera pixels.
• Learning: The agent improves its behavior from data or experience. In supervised learning, a model adjusts parameters using labelled examples such as images paired with object names.
• Reasoning: The system draws conclusions from facts, rules, or probabilities; a medical system may combine symptoms and test results to estimate possible diagnoses.
• Problem solving: The agent searches among possible actions to reach a goal. A route planner compares paths and selects one that minimizes a criterion such as travel time.
• Language ability: Natural-language systems interpret or generate human language, including questions, commands, summaries, and translations.
• Decision-making: The system selects an action under constraints and uncertainty, such as deciding whether a detected object requires braking.
• Adaptability: Intelligent behavior remains useful when the exact situation was not included in the original instructions or training data.

B. Concept of intelligence

The concept of intelligence in AI is operational: a system is considered intelligent when it performs a task appropriately and flexibly according to defined criteria.

• Agent perspective: An agent observes a state, chooses an action, and receives a result. A robot vacuum observes obstacles and selects movement while attempting to clean efficiently.
• Rational behavior: A rational agent chooses the action expected to maximize its performance measure given its available information.
• Symbolic and statistical forms: Expert systems may manipulate rules such as “if fever and cough, consider infection,” while neural networks learn numerical relationships from examples.
• Human comparison: Human intelligence is broad, embodied, social, and often conscious; machine intelligence is usually specialized, engineered, and measured on particular tasks.
• Limit of the definition: High task performance does not by itself prove understanding, self-awareness, emotions, or general reasoning.

III. Differences between Weak AI and Strong AI — Two Views of Machine Intelligence

Weak AI and strong AI distinguish present task-oriented systems from a proposed form of genuinely general machine intelligence. The distinction concerns capability and, in some interpretations, the nature of understanding.

A. Differences between weak AI and strong AI

Weak AI, also called narrow AI, is designed for a specific task or limited range of tasks. Strong AI, often associated with artificial general intelligence, would possess flexible, human-like intelligence across many domains.

• 1. Weak AI — specialized performance:

  • Purpose: It performs a defined function, such as spam filtering, speech recognition, recommendation, or image classification.
  • Method: It may use rules, machine learning, deep neural networks, or a hybrid architecture.
  • Example: A chess program can evaluate moves and defeat expert players, but it does not automatically understand cooking, politics, or physical navigation.
  • Current status: Most deployed AI systems, including voice assistants and fraud detectors, are weak AI even when their outputs appear sophisticated.

• 2. Strong AI — general capability:

  • Purpose: It would transfer knowledge across unrelated tasks, learn efficiently, reason in unfamiliar situations, and possibly understand its own mental states.
  • Expected properties: General problem solving, common-sense reasoning, long-term learning, language competence, and flexible adaptation.
  • Status: Strong AI remains a research goal rather than an established, universally accepted technology.
  • Philosophical issue: Some definitions require consciousness or genuine understanding, while others define general intelligence only by broad behavioral competence.

• Key contrast: Weak AI simulates intelligent performance in a bounded domain; strong AI would integrate abilities across domains.

• Transferability: A narrow model trained to detect tumors cannot automatically plan a spacecraft mission without new data, training, and system design.

• Evaluation: Weak AI is judged by measurable task outcomes; strong AI would require tests of generalization, reasoning, learning, and possibly consciousness.

• Misleading appearance: Fluent language or human-like interaction does not necessarily demonstrate general intelligence, because a system may generate outputs from learned patterns without human-style comprehension.

B. Applications and limitations

The weak–strong distinction helps explain both the usefulness and the limits of modern AI.

• Practical value: Narrow systems can outperform humans in restricted tasks, such as recognizing retinal disease patterns or detecting unusual credit-card transactions.
• Dependence on conditions: Performance may decline when data differ from training conditions, a problem known as distribution shift.
• Lack of common sense: A system may identify objects accurately but fail to understand physical, social, or causal implications.
• Safety concern: Errors in high-impact contexts require human oversight, testing, transparency, fallback procedures, and accountability.
• Generalization requirement: Strong AI would need to apply principles learned in one situation to new situations rather than merely repeat familiar statistical patterns.

IV. History of AI from beginnings to present day — Major Stages

The history of AI from beginnings to present day includes philosophical ideas about reasoning, the development of digital computers, symbolic programs, periods of reduced investment, and the recent growth of machine learning and generative systems.

A. History of AI from beginnings to present day

AI developed through several overlapping stages rather than through one continuous invention.

• Early intellectual foundations:

  • Logic and computation: Aristotle’s formal logic, later mathematical logic, and the work of Alan Turing helped establish the idea that reasoning could be represented as symbolic operations.
  • Turing’s contribution: In 1950, Turing discussed whether machines could exhibit intelligent behavior and proposed an imitation-based conversational test.
  • Neural inspiration: McCulloch and Pitts described a mathematical model of artificial neurons in 1943, linking neural activity with computation.

• Birth of the field, 1950s–1960s:

  • Dartmouth workshop, 1956: John McCarthy, Marvin Minsky, Claude Shannon, and others helped establish AI as a named academic field.
  • Early programs: Logic Theorist (1956) demonstrated automated theorem proving, while ELIZA (1966) simulated conversation through pattern matching.
  • Dominant approach: Symbolic AI represented knowledge using symbols, rules, search procedures, and explicit reasoning.

• Knowledge-based systems and limitations, 1970s–1980s:

  • Expert systems: Programs such as MYCIN used domain rules to support medical diagnosis; their knowledge was encoded by specialists.
  • Bottleneck: Manually acquiring and maintaining large rule sets was expensive and difficult.
  • AI winters: Reduced expectations, weak results, and limited computing resources led to periods of decreased funding and enthusiasm, especially in the 1970s and late 1980s.

• Machine learning expansion, 1990s–2000s:

  • Statistical methods: Systems increasingly learned relationships from examples rather than relying solely on hand-written rules.
  • Landmark event: IBM’s Deep Blue defeated world chess champion Garry Kasparov in 1997 through extensive search and evaluation.
  • Data and hardware: The internet generated large datasets, while improved processors enabled more complex models.

• Deep learning and present day, 2010s–2020s:

  • Image recognition: AlexNet’s 2012 ImageNet performance demonstrated the power of deep convolutional neural networks trained with GPUs.
  • Reinforcement learning: AlphaGo defeated Go champion Lee Sedol in 2016 by combining neural networks with search and self-play.
  • Transformers and generative AI: Transformer-based models advanced translation, language generation, image production, coding assistance, and multimodal interaction.
  • Current direction: AI development now emphasizes foundation models, responsible deployment, efficiency, explainability, privacy, bias reduction, and human–AI collaboration.

B. Significance and limitations

Each historical stage expanded AI’s capabilities while revealing new technical and social challenges.

• Changing representation: AI moved from hand-coded symbols toward learned numerical representations, although modern systems often combine both.
• Performance drivers: Progress has depended on algorithms, data, computing power, benchmarks, and practical access to deployment environments.
• Persistent challenges: Robust reasoning, factual reliability, fairness, interpretability, energy consumption, privacy, and accountability remain unresolved.
• Historical lesson: Claims of imminent human-level intelligence have repeatedly exceeded actual performance, so evaluation must rely on demonstrated capabilities rather than publicity.

V. Examples of Today’s Real-World Applications of AI — Applied Intelligent Systems

Today’s AI applications combine perception, prediction, language processing, optimization, and decision support. In most cases, they assist people or automate a limited function rather than exhibit general intelligence.

A. Examples of today's real-world applications of AI

AI is used across sectors whenever large amounts of data, repeated decisions, or complex patterns make computational assistance valuable.

• Healthcare: Models analyze medical images, estimate patient risk, help discover drugs, and support clinical documentation.
  • Concrete use: An imaging model can flag a suspicious region in an X-ray for review by a radiologist; it supports diagnosis but does not replace clinical responsibility.
• Finance: Banks use AI for fraud detection, credit-risk estimation, algorithmic trading, and customer-service chatbots.
  • Concrete use: A transaction with an unusual location, amount, and timing may receive a high fraud-risk score for further verification.
• Transportation: AI supports route planning, traffic prediction, driver-assistance systems, warehouse robots, and autonomous vehicles.
  • Concrete use: A navigation system combines map data and live traffic estimates to select a route with lower predicted travel time.
• Communication and language: Translation, speech-to-text, text prediction, search ranking, summarization, and generative assistants rely on language models.
  • Concrete use: Speech-recognition software converts an audio signal into written words for captions or dictation.
• Retail and entertainment: Recommendation engines personalize products, films, music, and advertisements using interaction histories and similarity patterns.
  • Concrete use: A streaming service ranks titles based on viewing history, content characteristics, and behavior of users with similar preferences.
• Manufacturing: Computer vision detects defects, while predictive maintenance models estimate when equipment may fail.
  • Concrete use: Abnormal vibration readings can trigger maintenance before a production-line motor breaks.
• Agriculture and environment: AI analyzes satellite imagery, soil readings, weather data, and crop images to improve irrigation and identify disease.
• Cybersecurity: Systems detect unusual network activity, classify malware, and prioritize alerts for security analysts.
• Education: Adaptive learning platforms recommend exercises, provide automated feedback, and identify topics requiring additional instruction.

B. Applications and limitations

The value of an AI application depends on the quality of its data, the consequences of error, and the way humans supervise its decisions.

• Accuracy is contextual: A 95% accurate model may be useful for low-risk recommendations but unsafe for a critical medical or aviation decision.
• Bias can be learned: If historical data reflect unequal treatment, a predictive system may reproduce or intensify that pattern.
• Privacy matters: Training or operating systems on medical records, faces, voices, or location histories requires lawful handling and appropriate safeguards.
• Human oversight: In high-impact settings, people should be able to review evidence, challenge outputs, and override automated recommendations.
• Security and reliability: Adversarial inputs, data poisoning, system failures, and fabricated outputs can produce harmful results.
• Social impact: AI can increase productivity and accessibility, but it may also change employment patterns, spread misinformation, or concentrate technological power.
• Responsible deployment: Effective use requires clear objectives, representative testing, monitoring after release, documentation, and defined responsibility for mistakes.