Week 1 Lecture 2: Different Types of Learning

Week 1 Lecture 2: Different Types of Learning

• Supervised Learning
  • Has (X, Y) given as data, where X was input, Y was output, and model tries to find out Y for a new X and compare with the given Y
  • Give a label to X (ie find Y)

• UnSupervised Learning
  • Only X is given
  • Given X, Cluster or Summarise them, ie organise them into meaningful groups
• Reinforcement Learning
  • Given an “agent” Determine what to do based on rewards and punishments
  • Agent takes an action which impacts the enviornment, based on which it is rewarded ( rewards can be [-1, 0, 1])
  • The agent tries to optimise immediate rewards
• SemiSupervised Learning //Not defined in Andrew Ng’s course, who put it in supervised learning only
  • Combo
  • Given labelled training data, and a even larger unlabelled data, come up with algo to process the unlabelled data
  • Most common for now

Supervised Learning

We have a set of input features given \( X_1, X_2, … , X_n \), with respect to which the instances are described. We also have a target feature Y.

So the values of \( Y_1\) are given. Along with this a test instance is given, containing only \( X_1\).

• If Y is discrete valued, [eg Will it rain, or not rain] it is known as Classification.
• If Y is continious valued, [eg Given a location, predict price of house per sq inches], it is known as Regression.
  • Most common example - linear regression

Features

Individual observations are analysed into a set of quantifiable properties which are called features eg Type of blood group, or Ordinal [small, medium or large] etc

To Summarise Supervised Learning

Classification Learning

• Task T is given
  • Input: Set of Instances \( d_1, d_2, … , d_n \)
    • Each instance has a set of features
    • So d can be represented as a vector \vec d = < \( X_1, X_2, … , X_n \)>
  • Output: A set of predictions as one of a fixed set of constant values
• Performance Metric P
  • Probability of a prediction being wrong
• Experience E
  • It is the data, ie a given set of correctly labelled examples (X,Y), coming from a fixed distribution (ideally)

TLDR

• Features: Distinct traits that is used to describe a particular instance quantitatively
• Feature Vector: n - dimentional vector of numerical features
• Instance space X: Set of all possible objects describeable by features
• Example (x, y): Instance x with label f(x) = y
• Concept c: Subset of objects from X [eg given a group of images, those images that are of a bicycle are under a concept c that contains images of bicycles]
• Target Function: Maps each instance x ε X to target label y ε Y
• Training Data: Collection of examples observed by learning algorithm, used to potentially predict relationships [in the Computer Science sense ;) ]