Connor's Notes

⬅️ Back to portfolio

COSC 3P71: Week 1 Notes

4 min read

Note

These notes are taken after the fact based on the slides. Anything mentioned only in lecture is not included. Covers the entire Week 1 AI introduction slideshow.

Introduction to Artificial Intelligence

What is AI?

  • Study of building intelligent entities: machines that act effectively & safely across diverse novel situations
  • Fastest-growing field; wide applications; still much room for contributions
  • AI spans:
    • General: learning, reasoning, perception
    • Specific: chess, theorem proving, self-driving cars, medical diagnosis
  • Universal field: relevant to any intellectual task

AI vs Machine Learning

  • Machine learning (ML): subfield of AI that improves performance from experience
  • Confusion in public use — ML is only part of AI
  • AI systems may use ML, but not all do
  • Two views:
    • Human-centered: AI as human-like intelligence
    • Rationalist: AI as “doing the right thing” (rationality)

Russell & Norvig advocate: AI = acting rationally

Four Categories of AI

AI can be classified by two dimensions: human vs rational and thought vs behavior:

HumanRational
ThinkThink like humans (cognitive modeling)Think rationally (laws of thought)
ActAct like humans (Turing test)Act rationally (rational agent)

Acting Humanly – The Turing Test

  • Turing (1950): “Can machines think?”
  • Proposed the Imitation Game:
    • If a computer can fool a human interrogator into thinking it is human, it demonstrates intelligence
  • Capabilities needed to pass:
    • Natural language processing
    • Knowledge representation
    • Automated reasoning
    • Machine learning
  • Total Turing test adds:
    • Computer vision
    • Robotics

Example: ELIZA chatbot — worked via simple syntactic tricks, not deep understanding.

Thinking Humanly – Cognitive Modeling

  • To claim a program thinks like a human, we must know how humans think:
    • Introspection (self-observation)
    • Psychological experiments
    • Brain imaging
  • Requires scientific theories of internal brain activity
  • Validated through prediction + testing or neurological evidence
  • Overlaps with cognitive science and cognitive neuroscience

Example: ML + brain imaging used to approximate “mind reading”

Thinking Rationally – Laws of Thought

  • Originates with Aristotle (“right thinking”)
  • Formal logic → rules of reasoning
  • Example: All men are mortal; Socrates is a man → Socrates is mortal
  • Problems:
    • Not all intelligent behavior involves logical deliberation
    • Logic requires certainty about the world (rare)
    • Probability theory helps with uncertainty
  • Rational thought ≠ rational behavior — need theory of rational action

Acting Rationally – Rational Agents

  • Rational behavior: doing the right thing to maximize goal achievement given available info
  • Not always about “thinking” (reflexes can be rational if optimal)
  • Rational agent:
    • Acts autonomously
    • Perceives environment
    • Persists over time
    • Adapts to change
    • Pursues goals
  • AI research goal = design rational agents
flowchart TD
    P[Percepts / History] --> F[f: P* → A]
    F --> A[Actions]

Limited Rationality

  • Perfect rationality = impossible (computational limits)
  • Instead: bounded rationality
    • Act “well enough” under time/knowledge constraints
  • Still, perfect rationality serves as a useful benchmark for theory

Beneficial Machines & Value Alignment

  • Standard model assumes objectives are fully specified
  • Works in artificial tasks (chess, shortest path)
  • But in real-world tasks, defining objectives is hard:
    • Example: self-driving cars
      • Goal: reach destination safely
      • Perfect safety → never leave garage
      • Must balance progress vs. risk
  • Value alignment problem: objectives given to AI must match true human value
  • Risks:
    • Misaligned AI might pursue objectives dangerously (e.g., bribe opponent in chess if “winning” is sole goal)
    • We want machines that are cautious, ask permission, defer to humans

Other Definitions of AI

  • Dean, Allen & Aloimonos: flexible programs responding productively in unanticipated situations
  • Winston: computations that perceive, reason, and act effectively in uncertain environments

Goals of AI

  • Engineering: solve real-world problems with knowledge & reasoning
    • Focus on higher-level design & intelligent software entities
  • Science: use computers to study intelligence itself
    • Test theories of human intelligence by implementing them in code

Perspectives on AI

  • Computer science: building theories/programs to solve problems
  • Cognitive science: simulate neurology and human cognition
  • Psychology: human intelligence studies
  • Philosophy: reasoning about perception, learning, memory

Splinter Fields of AI

  • Computer vision
  • Theorem proving / symbolic computation
  • Logic programming
  • Natural language understanding
  • Robotics
  • Data mining
  • Machine learning
  • Neural networks
  • Evolutionary computation/robotics
  • Swarm intelligence
  • Deep learning
  • Reinforcement learning
  • Large language models

Evolution of AI

  • AI is active, evolving, with conferences & journals
  • Phenomenon: once a problem is solved, it often “leaves AI” and becomes mainstream computer science
    • Examples: chess playing, OOP, theorem proving, pattern recognition

Advantages of Implementing Intelligence on Computers

  1. Problem-solving via computation
    • Links to tractability, complexity, PL paradigms
  2. Precision
    • Programs must be unambiguous
  3. Measurement
    • Enables empirical analysis
  4. Computers as guinea pigs
    • Ethical way to experiment with “minds”

State of the Art

  • Current frontier: covered later in class
  • Focus areas include:
    • Deep learning
    • Reinforcement learning
    • Large-scale perception & reasoning systems

Graph View