Machine Learning Fundamentals

Outline of the Machine Learning Course

• Course covers a wide range of topics from abstract theory to practical applications
• Progression starts with building concepts and theory, then moves to practical aspects
• Some topics intentionally placed earlier to provide tools for testing theoretical ideas

What is Learning?

• Machine learning is broad, covering both abstract theory and practical rules of thumb
• Essence of machine learning has 3 key components:
  • Pattern exists in the data
  • Data is available to learn from
  • Cannot be fully described mathematically

Movie Rating Example

• Netflix offered $1 million prize for 10% improvement in movie recommendation system

• Small improvement can have huge business impact if it makes recommendations more accurate

• To rate a movie:

  • Describe viewer as vector of factors (e.g. liking comedy, action, etc.)
  • Describe movie as vector of content factors
  • Compare viewer and movie vectors to predict rating

• This manual approach is not true machine learning

Learning Approach

• Machine learning reverses the manual process

  • Start with random viewer and movie vectors
  • Iteratively adjust vectors to match known ratings
  • Vectors become meaningful after processing many ratings

• Similar to credit approval problem:

  • Input is applicant information (age, income, etc.)
  • Output is credit decision (+1 approve, -1 deny)
  • Target function mapping input to output is unknown
  • Use historical data to learn the target function

• Key components:

  • Input data (applicant information)
  • Output (credit decision)
  • Target function (ideal credit approval formula)
  • Use historical data to learn the target function### Background

• The goal is to learn a function that approximates the unknown target function (f) using a given set of data

• The final hypothesis (g) is the function learned by the learning algorithm that aims to approximate f

• The hypothesis set (H) is the set of candidate functions the learning algorithm can choose from

Hypothesis Set

• The hypothesis set (H) contains the candidate functions (h) that have the same functional form
• Each hypothesis h is a function that takes input x and produces an output using parameters w and a threshold
• Adding an artificial input x0 = 1 simplifies the hypothesis function to a dot product of w and x

Perceptron Learning Algorithm

• The perceptron learning algorithm takes the training data and attempts to find the correct weight vector w
• If a data point is misclassified, the algorithm updates w by adding/subtracting the misclassified data point x
• This update rule moves the decision boundary to correctly classify the misclassified point
• The intuition is that the algorithm keeps adjusting w to improve classification of the data

Limitations

• The perceptron learning rule is not a "crazy" rule, but its viability and performance still needs to be evaluated
• The iterative updates of the perceptron algorithm can move the decision boundary to correctly classify the data### Supervised Learning
• Supervised learning involves training a system with data that has clearly labeled outputs, e.g. customer data with their actual credit behavior
• This allows the system to learn the relationship between inputs and outputs, and then apply it to classify new, unlabeled data
• Example: Coin recognition - use size and mass measurements of coins (inputs) with their known denominations (outputs) to train a system to classify new coins

Unsupervised Learning

• In unsupervised learning, the training data has no labeled outputs
• The system must find patterns and structure in the input data on its own
• Example: Coin data with no denomination labels - the system can still cluster the coins into groups based on their measurements, even though it doesn't know what the groups represent
• Unsupervised learning allows discovering higher-level representations of the data, like language models from listening to speech

Reinforcement Learning

• Reinforcement learning involves an agent interacting with an environment and receiving feedback (rewards/penalties) on its actions
• The agent learns an optimal policy by trial-and-error, adjusting its actions based on the feedback
• Example: Learning to play Backgammon - the agent tries different moves, gets feedback on wins/losses, and gradually learns the best strategy through many iterations
• Reinforcement learning is attractive when the target function cannot be easily modeled, as the agent can learn complex behaviors through experience

Machine Learning Course Outline

• Covers a wide range of topics from abstract theory to practical applications
• Progression starts with building concepts and theory, then moves to practical aspects

Key Components of Machine Learning

• Pattern exists in the data
• Data is available to learn from
• Cannot be fully described mathematically

Movie Rating Example

• Manual approach: describe viewer and movie as vectors of factors, compare to predict rating
• Not true machine learning, as no learning from data occurs

Machine Learning Approach

• Reverses the manual process: start with random vectors, iteratively adjust to match known ratings
• Vectors become meaningful after processing many ratings

Learning Approach Analogy

• Similar to credit approval problem: input is applicant information, output is credit decision
• Target function mapping input to output is unknown, use historical data to learn target function

Background

• Goal is to learn a function that approximates the unknown target function using given data
• Final hypothesis is the function learned by the learning algorithm that aims to approximate target function
• Hypothesis set is the set of candidate functions the learning algorithm can choose from

Hypothesis Set

• Contains candidate functions with the same functional form
• Each hypothesis is a function that takes input x and produces an output using parameters w and a threshold
• Adding an artificial input x0 = 1 simplifies the hypothesis function to a dot product of w and x

Perceptron Learning Algorithm

• Takes training data and attempts to find the correct weight vector w
• If a data point is misclassified, updates w by adding/subtracting the misclassified data point x
• Update rule moves the decision boundary to correctly classify the misclassified point

Limitations

• Perceptron learning rule is not guaranteed to perform well, needs to be evaluated
• Iterative updates can move the decision boundary to correctly classify the data

Supervised Learning

• Involves training a system with data that has clearly labeled outputs
• Allows the system to learn the relationship between inputs and outputs, then apply it to classify new data
• Example: Coin recognition with labeled denominations

Unsupervised Learning

• Training data has no labeled outputs
• System must find patterns and structure in the input data on its own
• Example: Coin data with no denomination labels, system can still cluster coins into groups based on measurements

Explore the basics of machine learning, covering theory and practical applications. Learn about the key components of machine learning and how to apply them.