Date: August 24, 2025

Topic: Linear Classifiers

Recall

Supervised learning requires examples and labelled data in the dataset.

In unsupervised learning, we only have examples and thus can only learn a probabilistic model.

In reinforcement learning, the supervision comes from a reward signal, which may be late in the agent’s life.

Notes

Types of ML

Supervised Learning

• Train Input: {$X, Y$}
• Learning Output: $f:X \rightarrow Y$ or $P(y|x)$
  • Inputs $x_i$ and $y_i$ are represented as vectors
• Each example has a ground truth label from a human annotator
• This dataset is then used to train the model

Unsupervised Learning

• We only have examples and no labels, ie Train Input: {$X$}
• Can only learn a probabilistic model, Learning Output: $P(x)$
• e.g., Clustering, density estimation, etc

Reinforcement Learning

• Agent acts in the world and can perform various actions affecting the environment
• Environment then changes its state based on the action, and gives the agent a reward
  • Supervision here is through a reward, not label
  • No supervision from human on what action to take, only have reward which can come late in the agent’s life

Non-parametric models do not model a function explicitly.

Parametric models explicitly model the function where $W$ are the parameters.

Supervised Learning

Non-parametric Model

• No explicit model for the function (e.g., kNN, Decision Trees)
• In above kNN, we just store the examples and during test time, find the nearest training sample

Parametric Model

• Explicitly models the function $f:X \rightarrow Y$ as a parametrized function $f(x,W)=y$, where $W$ are the parameters
• E.g., logistic regression, classification and neural networks