Machine Learning: Foundations & Models

Source: Unit 1 §1

What machine learning is

In our world we have humans and computers. Humans learn from experience; computers traditionally follow instructions. The question that starts the subject is whether a computer can learn from experience too, and the answer is yes, provided you accept that a computer's word for "experience" is data.

The whole premise of ML in one line: swap "experience" for "data" and the learner can be a machine.

The two broad phases

Training builds the model from data; prediction serves it on data the model has never seen.

Phase	What happens
Training	Use data to create and fine-tune a predictive model.
Prediction	Use the trained model to serve predictions on previously unseen data and answer queries.

The seven steps to machine learning

→StepsThe ML pipeline, in order

1. Gathering data - collect raw data.
2. Data preparation - clean, format, split.
3. Choosing a model - pick an algorithm.
4. Training - fit the model to the data.
5. Evaluation - test on unseen data.
6. Parameter tuning - improve performance.
7. Prediction - deploy and answer queries.

◎Exam cueMemory hook

Gather → Prepare → Choose → Train → Evaluate → Tune → Predict.

Formal definitions

Herbert Simon: "Learning is any process by which a system improves performance from experience."

Tom Mitchell (1998): Machine Learning is the study of algorithms that improve their performance P at some task T with experience E. A well-defined learning task is given by the triple ⟨P, T, E⟩.

≡FactsCheckers, as a ⟨T, P, E⟩ triple

• T (task) - playing checkers.
• P (performance measure) - percentage of games won.
• E (experience) - playing practice games against itself.

⚠GotchaA task is not well defined until all three are named

"Learn to play checkers" is not a learning task, because Mitchell's definition is a triple. Without P you cannot tell whether learning happened, and without E you have not said what the learner is allowed to see.

When machine learning is the right tool

≡FactsFour situations that call for learning

• Human expertise does not exist - navigating on Mars.
• Humans cannot explain their expertise - speech recognition.
• Models must be customised - personalised medicine.
• Models are based on huge amounts of data - genomics.

The classic hard task is recognising a handwritten 2. Writing down the rules for what makes a 2 a 2 is very hard; learning it from examples is easy. That gap is the whole argument for the field.

Category of task	Examples
Recognising patterns	facial identities and expressions, handwritten or spoken words, medical images
Generating patterns	generating images or motion sequences
Prediction	future stock prices, currency exchange rates

The three broad categories

The three broad categories, separated by what the training signal is: labels, nothing, or rewards.

Type	Given	Goal
Supervised (inductive)	training data + desired outputs (labels)	learn a mapping input → output
Unsupervised	training data without labels	find hidden structure, e.g. clustering
Reinforcement	rewards from a sequence of actions	learn a policy, state → action

Supervised learning

Given pairs $(x_1,y_1), (x_2,y_2), \ldots, (x_n,y_n)$, learn a function $f(x)$ that predicts $y$ from $x$. The type of $y$ splits the family in two.

⇄CompareRegression vs classification

Regressiony is real-valued (continuous). Predict the price of a house.

Classificationy is categorical (discrete). Decide spam or not-spam.

Unsupervised learning

Given $x_1, x_2, \ldots, x_n$ without labels, output the hidden structure behind the $x$'s. The canonical instance is clustering: in genomics, group individuals by genetic similarity when nobody has told you which groups exist.

Reinforcement learning

Given a sequence of states and actions with (possibly delayed) rewards, output a policy - a mapping from states to actions that tells the agent what to do in each state.

The RL loop at discrete time steps: the agent acts, the environment answers with a new state and a (possibly delayed) reward.

Designing a learning system

→StepsFour design choices

1. Choose the training experience - what data, and how it is gathered.
2. Choose exactly what is to be learned - the target function.
3. Choose how to represent the target function.
4. Choose a learning algorithm to infer the target function from the experience.

The OSI analogy for the ML pipeline

The slides map the ML pipeline onto the seven-layer OSI stack to show it as a layered flow, with raw data at the bottom and decisions at the top.

OSI layer	ML pipeline stage	Level
Application	Application / decision making	high
Presentation	Decision making
Session	Classification
Transport	Feature extraction
Network	Segmentation
Data link	Processing
Physical	Raw input data	low

Takeaway: ML transforms raw input data → segmentation → feature extraction → classification → decision, layer by layer.