1 - ML Overview

What is machine learning?

The study of computer programs (algorithms) that learn by example and can generalize from existing examples of a task.

Classical programming: inputs and outputs

Data + Rules → Answers

Machine learning: inputs and outputs

Data + Answers → Rules (the model learns the rules)

What is supervised learning?

Training a model using labeled data to learn rules that predict labels on unseen data.

What is unsupervised learning?

Finding useful structure or knowledge in data when no labels are available.

What is classification?

Predicting a class label, which is a choice from a predefined list of possibilities.

What is regression?

Predicting a continuous (real-valued) number.

Classification or regression: predicting whether an email is spam?

Classification (the output is a category).

Classification or regression: predicting a house's sale price?

Regression (the output is a continuous number).

Examples of unsupervised learning

Clustering similar users, social network analysis, market segmentation, and detecting abnormal server access patterns (outlier detection).

What is clustering?

An unsupervised task that groups similar data points together without using labels.

Supervised vs. unsupervised: the key difference

Supervised uses labeled data to predict labels; unsupervised has no labels and looks for structure in the data.

What is structured data?

Data organized into a fixed format, such as rows and columns in a table.

What is unstructured data?

Data without a predefined format, such as text, images, audio, and video.

Which type of data is deep learning especially good at?

Unstructured data (images, text, audio).

Traditional ML workflow

Input → manual feature extraction → classification → output.

Deep learning workflow

Input → feature extraction + classification learned together → output (features are learned, not hand-designed).

What is deep learning?

Hierarchical feature learning: a hierarchy of representations with increasing levels of abstraction.

Main difference between traditional ML and deep learning

Traditional ML needs manually engineered features; deep learning learns features automatically from raw data.

How do AI, ML, and DL relate?

DL is a subset of ML, and ML is a subset of AI. CV and NLP are application areas that overlap with them.

Why are there so many ML algorithms?

Different problems, data types, and goals suit different algorithms, so no single algorithm is best for everything.

What is model generalization?

A model's ability to make good predictions on new, unseen data, not just the data it was trained on.

Why split data into training, validation, and test sets?

To train the model, tune it, and then get an honest estimate of its performance on unseen data.

Purpose of the training set

Used to fit the model's parameters (learn the rules).

Purpose of the validation set

Used to tune the model and choose its complexity or hyperparameters.

Purpose of the test set

Used once at the end to estimate how the final model performs on unseen data.

Why not use the test set to choose the model?

It leaks information from the test set into your choices, so the test score becomes an overly optimistic estimate of real performance.

What is underfitting?

The model is too simple to capture the pattern: training error is high and validation error is high.

What is overfitting?

The model memorizes the training data (including noise): training error is low but validation error is high.

How to fix overfitting?

Get more data, reduce model complexity, or add regularization.

How to fix underfitting?

Increase model complexity, or reduce or remove regularization. More data will not help.

Learning curve signs of underfitting

Training and validation errors converge to similar values, but both are poor (high). More data will not help.

Learning curve signs of overfitting

A big gap between low training error and high validation error. More data will probably help.

Model complexity trade-off

Too simple → underfitting; too complex → overfitting. The goal is the complexity that gives the lowest validation error.