Machine Learning · Module 1

What Machine Learning Is

Take each lesson, answer its quiz, then do its practice. After the last lesson come the module quiz, the project and the interview questions.

Lessons

1.1What Is Machine Learning?Tell a program written from rules apart from a program learned from examples, and say what "learning" concretely means: parameters chosen to fit data.
1.2The Kinds of Machine LearningFrom the question being asked, tell whether a problem is regression, classification, clustering or something else.
1.3The Machine Learning WorkflowKnow the steps every project goes through, and what goes wrong when one is skipped.
1.4Samples, Features and LabelsTurn a real table into `X` and `y`, and name each column's type.
1.5Where Data Comes FromSee how collection and labelling decide what a model can learn, before any algorithm runs.

Module material

Take these after the last lesson, in this order. Next on each page leads through them.

Glossary

Every term the module introduces, with the lesson that introduces it.

example: One item of data, with its label: a message and spam or ham. 1.1
label: The answer a program should give for an example. 1.1
rule-based program: A program that follows a rule a person wrote, such as "spam if it contains free". 1.1
accuracy: The fraction of answers that are right: (spam caught + ham passed) / all. 1.1
baseline: The simplest program that answers, which a model has to beat: here, "ham" for every message. 1.1
feature (reminder, Maths 00.1): A measured value of an example that the model uses: the digit count. 1.1
parameter: A number inside a model that is chosen from data: t. 1.1
model: The form of the rule, with its parameters left open: "spam when the digit count is at least t". 1.1
training (restated from Maths 00.1): Choosing a model's parameters to make the fewest errors on labelled examples. 1.1
training data: The labelled examples used in training: 4,459 messages. 1.1
held-aside data: Labelled examples kept out of training, used only to score the model: 1,115 messages. 1.1
inference: Running a trained model on examples it was not trained on. 1.1
machine learning: Choosing a program's parameters from examples, then using it on new examples. 1.1
supervised learning: Learning from examples whose labels are known. 1.2
regression: Supervised learning where the answer is a quantity, so being close counts. 1.2
classification: Supervised learning where the answer is one of a set of categories. 1.2
binary: Classification with two classes, such as spam or ham. 1.2
multi-class: Classification with more than two classes, such as the ten digits. 1.2
unsupervised learning: Finding structure in examples that have no labels. 1.2
clustering: Finding groups of similar examples (Module 12). 1.2
anomaly detection: Finding unusual examples (Module 13). 1.2
dimensionality reduction: Describing examples with fewer columns (Module 11). 1.2
semi-supervised learning: Learning from a few labelled and many unlabelled examples. 1.2
self-supervised learning: Learning from labels taken from the data itself, as a language model predicts the next word. 1.2
reinforcement learning: Learning which actions to take from rewards; outside this track. 1.2
workflow: The ordered steps of a machine learning project: frame, collect, split, explore, baseline, model, evaluate, iterate, deploy, monitor. 1.3
held-aside data (reminder, ML 1.1): Examples set aside before any model is built, used only to score it. 1.3
evaluation: Scoring a trained model on data it was not trained on. 1.3
iteration: One change to the model, then evaluating again. 1.3
memorisation: Storing the training examples' answers instead of learning a rule, as model M does. 1.3
deployment: Running a model on new data as part of a real system. 1.3
monitoring: Measuring a deployed model's score on new labelled data. 1.3
validation set (narration): Data kept apart for choosing between versions of a model, so the final test is used once (Module 4). 1.3
sample: One row of the table, one thing being described: a car. 1.4
label (target) (reminder, ML 1.1 and Maths 00.1): The column the model predicts: mpg. 1.4
feature: A column the model uses to predict the label. 1.4
feature matrix: The table of features, one row per sample, written X. 1.4
label vector: The labels, one per sample, written y. 1.4
shape: The number of rows, and of columns, of an array: (398, 7); a vector's has one number, (398,). 1.4
numerical: A column whose values are amounts; continuous when any value in a range is possible, discrete when it counts. 1.4
categorical: A column whose values name groups, even when they are stored as numbers. 1.4
time (narration): A column whose order matters and whose differences mean something, such as the model year. 1.4
text data: A column of free text. 1.4
ordinal: Categories with an order, such as a kitchen graded excellent to poor. 1.4
image data: Pixel values, one feature per pixel. 1.4
missing value: A cell with no recorded value, read by pandas as NaN. 1.4
feature space (reminder, Maths 2.1): The space with one dimension per feature, in which each sample is a point. 1.4
human labelling: People deciding the label of each example. 1.5
sampling frame (reminder, Stats 1.7): The set of units that could end up in a sample; here, four particular collections of messages. 1.5
selection bias (reminder, Stats 1.7): When the way rows are chosen makes the data differ from the population a claim is about. 1.5
survivorship bias (reminder, Stats 2.3): A summary built only from the cases that passed a selection step: the spam people complained about. 1.5
collection artifact: A pattern in the data made by how it was collected, not by what it records. 1.5
duplicate: A row whose content repeats another row. 1.5
data-quality check: A test of the data for problems, run before any model. 1.5
annotator: A person who labels examples. 1.5
label noise: Labels that are wrong. 1.5
train/serve mismatch: A difference between the data a model was trained on and the data it runs on. 1.5