notes.md

CS 4641 - Machine Learning

• What is machine learning?

  • The art, science, engineering, math, computing, of building systems that adapt and improve over time (typically in response to data)
  • Supervised
    • Neural Nets
    • Decision Trees
    • Function Approximation, given inputs and outputs, approximate a function that given an unknown input can guess an output
    • Inductive
      • Going from a series of datapoints to a general rule
      • "If you don't believe the sun is going to come up tomorrow, why are you here?"
    • in general, y=f(x), find f
  • Unsupervised
    • description or partition
    • efficient/compact way to describe the data
    • what is the best way
    • inductive bias in both
    • compression is a form of unsupervised learning
  • Reinforcement
    • you get results from actions
    • decision making
    • goal is to maximise "good" results from a series of actions

• "some of you have called me Cedric, you racist bastards"

• in machine learning, the single most important thing is the data, the algorithm is not important, at least not as important as the data

• "Chess is like pacman, for sufficiently large values of pacman"

• classification is function approximation, and thus deals with discrete values

  • you have discrete things that a classifier maps to
  • these are chosen based on aspects of the training data

• decision trees

  • powerful
  • can, without loss of generality, represent all boolean functions
  • 2^2^n decision trees
  • must be clever when building decision trees
  • best attribute is the most important, is the one that removes the most questions
  • entropy has to do with random
    • you need to reduce randomness to have effective decision trees
  • Restriction bias
    • the set of functions you're willing to use
  • shorter trees are better than longer
    • fewer nodes/smaller
  • high information gain is better than low
  • correctness
  • IM3 to pick a good one

• Overfitting is bad yo

  • don't
  • low error on training and high error on test
  • inferring about the noise instead of the underlying structure of the data

• Neural Nets

  • made up of perceptrons
    • a perceptrion is an output based on a set of weighted inputs and a linear transformation
  • can represent any arbitrary function
    • so picking the correct one is hard
    • use gradient descent algorithm to pick
      • gradient descent also has inductive bias
        • prefers systems with local maxima (perturbations cause no change)
        • correctness
        • prefers simpler functions