Tue April 23

1. Progress on projects?

2. Course evaluations.

3. Discuss AI wikipedia's article

• What is it?
  • (a) Have machines do what people can do. ("weak ai")
  • (b) Make a machine that is a person. ("strong ai")
  • (c) "AI is whatever hasn't been done yet." - Maloof
    • So ... what has been done?
    • reasoning (e.g. medical knowledge base)
    • natural language processing (e.g. text classification)
    • perception (sound ... Siri , vision ... captchas)
    • motion and manipulation (robots, cars)
    • learning (train a model with examples)
• For a given task, how we do it VS how machines (might) do it
  • Do submarines swim?
  • Experience so far is that with machines, statistical approaches win.
• Philosophy ...
• machine learning (e.g. kaggle.com )
  • ... has seen progress in the last few years
  • approach :
    • set of (inputs) => output
    • invent a "is this success" function
    • computer model : neural net, decision tree, equation, ... with unkown paremters
    • compute parameters by "training" - gradient descent or similar
    • Then given a new input, it gives an answer.
  • Simplest version of this : LINEAR REGRESSION!
    • A lot of the ideas from curve fitting apply here:
      • over fitting ... why that's bad and
      • gradient descent : heading toward a better solution in small steps
  • Examples ... will do some of these in more depth Thursday
    • optical character recognition
    • robot navigation with "liklihood" cloud & "evolution"
    • Google's AlphaGo :
    • this person does not exist ... "adverserial" approach