Back to Contents

Machine Learning NPTEL



Introduction

  • Tom Mitchell 1997 > said to learn from experience with respect to some class of tasks and a performance measure P, if [ the learner's] performance at tasks in the class, as measured by P, improves with experience

  • Requirements-

    • Class of tasks
    • Performance Measure
    • Experience

  • Paradigms of ML

    • Supervised Learning
      • Classification : Categorial Output
      • Regression : Continuous Output
    • Unsupervised Learning
      • Clustering : cohesive grouping
      • Association : frequent co-occurances

    • Reinforcement Learning

      • Learning Control
      Task Measure
      Classification error
      Regression error
      Clustering scatter/purity
      Associations support/confidence
      Reinforcement Learning cost/reward

  • Supervised Learning : Applications

    • Credit Card Fraud detection
    • Sentiment Analysis
    • Churn prediction
    • Medical Diagnosis

  • Unsupervised learning :

    • finding clusters -- bias => shape of cluster boundary in vector space
    • Association rule mining
    • Mininig Transactions

  • Reinforcement Learning :

    • Trial and error
    • self learning
    • feedback from environment
    • control system
    • Applications:
      • Game Playing
      • Autonomous agents
      • Adaptive Control -- quadrotor control
      • Combinatorial optimization -- VLSI placement
      • Intelligent Tutoring System



Proabability Theory

  • Properties of Set operations:

    • Commutativity
    • Associativity
    • Distributivity
    • DeMorgan's Law

  • Two events A and B are disjoint (or mutually exclusive) if A intersection B is empty set

  • Sigma algebra is a collection (F) of subsets of sample space (O) such that-

    • phi belongs to F
    • if A is in F then AC is also in F
    • If Ai for every i in N, then Uinfi=1 Ai in F

    A set A that belongs to F is called F measureable set (event)

  • power set is always sigma algebra

  • The concept of sigma algebra is powerful in the case where sample space (O) is uncountable

  • Probability Measure and Probability Space

    • A probability measure P on (O,F) is a function P:F->[0,1] satisfying: a. P(phi) = 0, P(O) = 1 b. if A1, A2 ... is a collection of pair-wise disjoint members of F, then:



      Back to Contents