Machine Learning Course - Lecture 1

Course Overview

• Lecturer: Yaser Abu-Mostafa
• Topics: The course covers a mix of mathematical theory and practical aspects of machine learning. Topics are color-coded to indicate this balance.
• Structure: Course follows a storyline:
  • What is learning?
  • Can we learn?
  • How to learn?
  • How to learn well?
  • Take-home lessons
• Exception: Third lecture is a practical topic included early for tools to test theoretical aspects.

Today's Lecture: The Learning Problem

• Logo: The course's logo is a technical figure to be discussed later.
• Outline:
  • Introduction to Machine Learning
  • Example: Movie ratings prediction
  • Mathematical formalization of the learning problem
  • First machine learning algorithm
  • Survey of types of learning
  • Puzzle to understand learning intricacies

Example: Movie Rating Prediction

• Problem: Predict how a viewer would rate a movie (e.g., Netflix).
• Components of learning problem:
  • Pattern Exists: Viewer ratings are consistent with other ratings and personal tastes.
  • Cannot Pin Down Mathematically: Need to learn from data as we cannot manually define a predictive function.
  • Data Availability: Essential for learning.
• Solution Approach:
  • Describe viewer/movie as vectors of factors (e.g., comedy, action, etc.).
  • Machine Learning Approach: Start with random factors and adjust based on data (ratings) until patterns emerge.

Components of Learning

• Applicant Information (Example: Credit Approval):
  • Input (x): Customer application data (e.g., age, salary)
  • Output (y): Approval decision (+1 or -1)
  • Target Function (f): Ideal unknown formula for approval.
  • Data: Examples from historical records.
  • Hypothesis (g): Approximation of f, derived from data.
• Learning Algorithm: Processes data to produce g.
• Hypothesis Set (H): Set of possible hypotheses from which g is chosen.

Perceptron Model: Example Hypothesis Set and Algorithm

• Input: Vector of customer attributes
• Hypothesis Set: Linear combination of attributes
• Learning Algorithm (Perceptron):
  • Start with random weights
  • Adjust weights to reduce misclassification
  • Guaranteed to converge if data is linearly separable
• Linear Inseparability: Techniques to handle this will be discussed.

Types of Learning

• Supervised Learning: Data includes input-output pairs. Focus of the course.
  • Example: Coin recognition
• Unsupervised Learning: Only inputs are provided; cluster finding.
  • Example: Data clustering without labels
• Reinforcement Learning: Learning through graded responses.
  • Example: Game playing (backgammon)

Learning Puzzle

• Puzzle: Given known examples, predict the output of an unknown function.
• Illustrative Point: Highlights the challenge of generalizing from finite data to unknown future data.

Q&A Notes

• Linearly Separable Data: Techniques like mapping and modifications (e.g., pocket algorithm) will be covered.
• High-dimensional Data: Computational challenges increase with dimensions.
• Pattern Detection: Learning feasible when a pattern exists, determined via theory.
• Bias and Sampling: Address sampling bias and model generalization.
• Types of Hypotheses: Finite and continuous hypothesis sets, generalization based on the size and complexity.
• Feedback Mechanisms: Validation and reinforcement learning as feedback methods.