Machine learning concepts

For A4.1.1 you need to describe each type of machine learning and a real-world use for it. Most types have a demo here.

• A4.1.1

  Supervised learning

  Learns from labelled examples, where each input comes with its known answer, then predicts the answer for new inputs. Used for classification (discrete labels, like marking an email as spam or not) and regression (continuous values, like predicting a house price).

  k-NN demo →
• A4.1.1

  Unsupervised learning

  Unsupervised learning finds structure in unlabelled data, such as clusters of similar items. Shops can use it to group customers with similar buying habits.

  k-means demo →
• A4.1.1

  Reinforcement learning

  In reinforcement learning, an agent learns from actions and the rewards it receives in an environment. It aims to maximise cumulative reward (reward over a sequence of actions). This approach is used in game-playing systems and robotics.

  grid-world demo →
• A4.1.1

  Deep learning

  Deep learning uses artificial neural networks with many layers. These learn features from data such as images and text. It is used in image recognition and language models.

  CNN demo →
• A4.1.1

  Transfer learning

  Transfer learning adapts a model trained on one task for a related task. Reusing learned features can reduce the data and training time needed. For example, an image model can be adapted to identify plant diseases by retraining its final layers.

For A4.1.2 you need to describe which hardware suits each situation where machine learning is used, and why.

• A4.1.2

  GPUs

  Graphics Processing Units (GPUs) perform many calculations in parallel. This suits the matrix calculations used to train neural networks.

  practise the GPU question →
• A4.1.2

  TPUs

  Tensor Processing Units (TPUs) are chips Google designed for neural-network calculations. They are specialised for efficient matrix operations, with less flexibility than general-purpose processors.

• A4.1.2

  ASICs

  Application-Specific Integrated Circuits (ASICs) have hardware designed for a particular purpose. They can offer high speed and low power use for that work. The hardware design is fixed after manufacture, but some ASICs can run different programs or models. A TPU is an ASIC designed for machine learning.

• A4.1.2

  FPGAs

  Field-Programmable Gate Arrays (FPGAs) have hardware that can be reconfigured after manufacture. They can accelerate suitable tasks while allowing the hardware design to change. They are useful for prototyping and for devices whose processing requirements may change.

• A4.1.2

  Cloud platforms and HPC centres

  Cloud platforms provide computing resources on demand. High-performance computing (HPC) centres provide powerful systems. Both can support large training tasks.

For A4.4, discuss the ethical implications of machine learning. Use the two activities below to practise considering different perspectives.

• A4.4.1

  Implications of ML systems

  The syllabus lists accountability, fairness, bias, consent, privacy, security and environmental impact. A model can be accurate overall and still treat one group unfairly, so high accuracy on its own does not show that a model is fair.

  accuracy-vs-fairness demo →
• A4.4.1

  Argue a real case

  You get better at ethical reasoning by practising it on real cases. Take a position on a real machine learning scenario and defend it, with CSBot arguing the other side.

  Ethics Court →
• A4.4.2

  Technology in daily life

  As AI, augmented reality (AR), virtual reality (VR) and quantum computing become part of everyday life, weigh their benefits against individual rights, privacy and equity. For each one, ask who benefits and who could be harmed.

The HL-only A4 topics that the demos do not cover directly.

• A4.2.1

  Data cleaning

  Real data is messy: missing values, duplicates, inconsistent formats, outliers. Cleaning it first matters because a model can only learn from the data it is given, errors and all.

• A4.2.2

  Feature selection

  Feature selection means choosing which inputs (features) the model uses. Filter methods rank features with a statistic before training. Wrapper methods train the model on different sets of features and keep the best set. Embedded methods choose features during training. Fewer, more useful features usually make a model faster and better at working on new data.

• A4.2.3

  Dimensionality reduction

  Dimensionality reduction uses fewer features while preserving useful information. It can reduce training time and overfitting. The syllabus requires the purpose and benefits, without the details of methods such as PCA and LDA.

• A4.3.2

  Decision trees

  A decision tree classifies an item by testing its features. Each result leads along a branch to a leaf that gives the predicted category. During training, the algorithm chooses splits that separate the classes. You can follow the path to explain a prediction. The syllabus includes decision trees and k-NN for classification.

• A4.3.3

  Hyperparameter tuning

  Hyperparameters are settings chosen outside the values learned during training, such as k, polynomial degree and learning rate. Tuning compares settings on a validation set using a suitable measure, such as accuracy, precision, recall or F1 score. The aim is to perform well on new data. Several sliders here change hyperparameters.

  metrics demo →
• A4.3.5

  Association rule learning

  Association rule learning finds relationships between attributes across records. Market basket analysis uses it to find items often bought together. Rules can take the form "if bread and butter, then milk". Support measures how often all these items occur together. Confidence measures how often milk occurs among purchases containing bread and butter.

• A4.3.10

  Model selection and comparison

  No single algorithm is best for every problem. Compare the candidate models fairly by testing them all on the same unseen data, using a metric that suits the task. Then choose, weighing the score against trade-offs such as speed and how easy the model is to explain.

  metrics demo →