Lecture 13: What is AI?/Neural Network Basics

What is AI?

Artificial Intelligence

• What even is it?

Bad news

• No such thing as AI
• Bye

Artificial

• "Made by people"?

Intelligence

• …?
• Uh oh

It gets worse

• Can we achieve human-level generalization and abstraction?
  • Can we do that without human-level error and cognitive hazards?
  • (Prejudice, change blindness, sloppy reasoning…)

Ethical issues

• This conception of AI has ethical problems
• Sure, displacing human labor…
  • Technical change without social change reinforces power structures

Really big ethical issues

• Worst of all: "hard AI" is a fantasy of slavery
  • The worker who never tires or strikes
  • The "lover" who can't say "no"
• "R.U.R." written in 1920
  • "Robota" means "forced (serf) labor" in Czech

AI's "utopia"

Within every dystopia, there's a little utopia —Margaret Atwood

Now what?

• So, to recap:
• No such thing as AI
• It's also impossible
• Even if there were and it were, it should be illegal
• But we have a way out:

Artificial what now?

The question of whether Machines Can Think… is about as relevant as the question of whether Submarines Can Swim. —Edsger W. Dijkstra

Automated Decision-Making

• Let's try it this way:
• Lowercase-ai systems make decisions about problems
  • Important: Who acts on them?
• Successful ai systems make "better" decisions?

Types of decisions

• I like to break it down into three types of decisions:
• Satisfaction
• Optimization
• Prediction

Types of problems

• Addressing three types of problems:
• Synthesis
• Dynamics
• Agent behavior

Russell & Norvig

• Here's another common breakdown

  • Source: "AI: A Modern Approach", Russell & Norvig

    	Human	Rational
    Thinking	Think like a human	Think rationally
    Acting	Act like a human	Act rationally

Symbolic vs Non-Symbolic

• A different breakdown
• Symbolic: Machines reason over (human readable) symbolic representations
  • E.g. deriving a correct world model by analyzing experts' processes
• Non-Symbolic: Machines reason over opaque,(usually) learned models
  • E.g. getting correct inputs and outputs by analyzing data
  • Almost purely learning correlations

State of the Field

What is happening in AI right now?

Challenges

• Perception
• Computer vision
• Natural language processing/understanding/synthesis
• Knowledge representation/common-sense
• Learning
• Reasoning & problem-solving
• Robotics
• …

Achievements

• Capturing spoken language as text is getting pretty okay
  • Is this the same as understanding spoken language?
• AI can speak and sing (easier in some languages, interestingly)
• Self-driving cars
  • Good on highways
  • Okay off-roading
  • Urban driving very hard
    • Why?

Self-driving cars

• Computer-assisted driving is pretty great already
• Anti-lock brakes
• Automatic braking
• Pedestrian detection
• Blind spot warnings
• Lane drift avoidance
• Smart cruise control
• …

Emotions

• It's generally not too hard for humans to read emotions
• It's super hard for computers
  • People have different faces
  • People emote with their whole body
  • People don't always say what they mean
• Limited success so far

Automated Reasoning

• Really good on isolated problems
  • Chess, Go, Starcraft, …
• Really hard in general

Robots

• Walking with two legs is hard
• "Let robots be robots"

Fundamental issue

• AI isn't just one thing or one technique
• Figuring out what approach to use is more than half the challenge!

Automated Decision-Making

• AI is most useful (to me) as a method of inquiry
  • "Build it to understand it"
• I'm an old-fashioned symbolic AI guy
  • Of course my research also incorporates modern tools
  • It's not good to be too dogmatic