Machine Learning

Motivation

• AI goal: replace human programming with "self-programming" (= predict appropriate behavior based on experience)

• The example: infants

  • language skills
  • motor skills
  • other behaviors

• Usual dichotomy:

  • Algorithmic/heuristic "tricks"

  • Simulate human behavior (infant brain)

ML and Systems

• Data flow in ML systems

  • Supervised: Train first, then use

  • Reinforcement: Train during use

• Data complexity

  • Boolean, discrete, continuous

  • This course mostly Boolean

• ML system evaluation

  • statistical significance

  • negative and positive examples

  • Type I (false-positive) and Type II (false-negative) error

  • Receiver Operating Characteristic (ROC) curves

Modes Of Learning

• "Discovery" learning and "generalization" learning

  • TD-Gammon

  • Relation between discovery learning and generalization learning

• Deductive learning: concluding things from principles

  • Theorem proving

  • Knowledge compilation

  • Deductive "KR-based" learning in some database

  • Basic idea: cache most general easily-derivable version of result in case similar query later

Inductive learning

• Supervised vs. reinforcement

• Training, validation (test), and classification data

• Overtraining and overfitting

• Ensemble learning: e.g. boosting