`Lecture 4 - Machine Learning`

What Is Machine Learning?

Learning = improving at a task after observing data.
An agent learns if its performance improves after making observations about the world.
When the agent is a machine (algorithm), it's called machine learning.

Example:

A weather-predicting algorithm first assumes 50% rain chance.

After observing data (humidity, temperature), it adjusts to make better predictions.

*This slide defines machine learning as when an agent—in this case, a computer or algorithm—improves task performance by making observations about the world. It builds a model from data, which acts both as a hypothesis about the world and a piece of software that solves problems.*

Why Not Just Program Everything Manually?

In some problems (like face recognition), programming every rule is impossible or inefficient.
Manually coding rules:
- Works only for a small number of known cases
- Requires huge effort to adapt to new situations
Machine Learning allows the model to generalize to new, unseen examples.
- Useful when:
  - The problem is complex
  - The rules aren’t fully known
  - You want the system to adapt to new data

🔍 Deductive vs Inductive Learning

Reminder, from Semester 1: Inductive learning means reasoning from specific examples to general rules, while deductive learning means reasoning from general rules to specific conclusions.

Deductive Learning	Inductive Learning (Machine Learning)
Uses rules to derive conclusions	Learns patterns from data
Requires prior knowledge	Requires only a dataset, but no prior knowledge
Conclusions are guaranteed (if rules are sound)	Conclusions are approximate, not guaranteed

This slide contrasts *deductive learning* (not ML) and *inductive learning* (ML). Deductive learning relies on *prior knowledge* and produces *guaranteed correct* conclusions if the rules are sound. In contrast, machine learning is *inductive*: it requires *no prior knowledge*, learns *directly from data*, and produces *approximate* conclusions without guaranteed correctness.

so, what does it look like?

Structure of Machine Learning Problems

Think of a matrix (table):

Rows = examples (data points)
Columns = features (inputs) and label (output)
- Inputs = attributes like day, subject, time
- Output = label like "attend: yes/no" or a number

This slide shows how a machine learning problem is structured: the *input* is a vector of *features* or *attributes*, and the *output* is either a *finite label* (*classification*) or a *numerical value* (*regression*). This setup defines how data is represented in ML and what kind of prediction task is being solved.

This slide shows a typical *machine learning dataset* with *input features* (subject, time, day) and a *finite output label* (yes/no), used for a *classification* task. Each row is an *example*, and the goal is to learn patterns to predict the output for new combinations of inputs.

Types of Output

Classification:
- Output is from a finite set (e.g., Yes/No, Cat/Dog)
- Task = assign a label to each input
Regression:
- Output is a number (e.g., "52 students will attend")
- Task = predict a continuous value

Example:

Input: Math class, Monday, Early

Output (classification): No

Output (regression): 18 students