CS221 Lecture 1: Introduction to Artificial Intelligence

Jul 2, 2024

CS221 Lecture 1: Introduction to Artificial Intelligence

Welcome and Introductions

  • Instructors: Percy and Dorsa
    • Percy teaches class with Dorsa, expert in robotics and robotic interactions
  • Teaching Team: CAs with interests in NLP, ML, data mining, computer vision, etc.
    • Head CA (PhD student): Natural language processing
    • Second-year MS students: ML, data mining, NLP
    • One MS student: Computer vision

Announcements

  • Weekly sections for review and advanced topics
    • Overview session for Python and probability this Thursday
  • First homework is posted on website, due next Tuesday at 11:00 PM via Gradescope
    • Gradescope code will be posted on Piazza

Importance of AI

  • No need to motivate importance; AI success stories are widespread
  • Superhuman AI in games (Jeopardy, Go, Dota 2, poker)
  • AI in practical applications (document reading, speech recognition, face recognition, medical imaging)
  • Policy impact and societal transformation debates around AI

Historical Background of AI

  1. 1956 Dartmouth Workshop: Foundational moment organized by John McCarthy, Marvin Minsky, Claude Shannon, etc.
    • Initial successes in game playing and theorem proving
    • Initial optimism, followed by the first AI winter due to limited compute and intrinsic problem complexity
  2. 1970s and 1980s: Focus on expert systems
    • Knowledge encoded as deterministic rules
    • Impact on industries but eventually led to the second AI winter due to over-promising and under-delivering
  3. 1943 and Neural Networks: McCulloch and Pitts' work on artificial neural networks
  4. Deep Learning Resurgence: Rediscovery of backpropagation, key breakthroughs like Yann LeCun's work and AlexNet
    • AlphaGo as a notable example of combining logic and neural networks

Two Traditions of AI

  • Logical AI (John McCarthy, etc.)
  • Neuroscience-Inspired AI (artificial neural networks)
  • Examples of synergy between these traditions: e.g., AlphaGo

AI as Agents vs. AI as Tools

  • AI as Agents: Recreating human intelligence capabilities (perception, motor skills, language, knowledge, reasoning, learning)
  • AI as Tools: Using AI to solve real-world problems
    • Examples: Satellite imagery for economic predictions, data center energy management, self-driving cars, fairness and bias issues

Course Overview

Modeling, Inference, and Learning Paradigm

  • Modeling: Simplifying the real world into a mathematically precise model
  • Inference: Asking questions about the model
  • Learning: Deriving model parameters from data

Key Course Topics

  1. Machine Learning: Data to Model
    • Focus on generalization from data to new experiences
  2. Reflex Models: No feedback loop (e.g., linear classifiers, deep neural networks)
  3. State-Based Models: Planning and actions in state spaces
    • Search Problems: Optimal path finding
    • MDPs: Dealing with randomness
    • Games: Adversarial environments
  4. Variable-Based Models: Solutions as variable assignments
    • Constraint Satisfaction Problems (CSP)
    • Bayesian Networks: Soft dependencies and noisy sensor readings
  5. Logic Systems: Deep reasoning with heterogeneous information (e.g., simple logic language system demo)

Course Logistics

  • Prerequisites: Programming, discrete math, probability
  • Coursework: Homeworks, exam, and project
    • Homeworks: 8 total, involve written and programming problems
    • Exam: Closed book with one page of notes, focuses on problem-solving
    • Project: Work in groups of three, various milestones
  • Policies: Gradescope submissions, 7 late days, use Piazza for communication
  • Honor Code: Collaborate but write independently, do not post assignments on GitHub, debugging permitted

Initial Technical Details

Optimization in AI

  • Discrete Optimization: Finding the best path or set of decisions (Dynamic Programming)
  • Continuous Optimization: Minimizing objective functions (Gradient Descent)

Example Problems

  1. Edit Distance
    • Use dynamic programming to find the minimum number of edits to transform one string into another
    • Optimal strategy via recurrence relations and memoization
  2. Linear Regression
    • Fit a line using least squares and gradient descent for continuous optimization
    • General method for solving machine learning problems via gradient descent

Next Steps: Begin with machine learning in the next lecture.