Subjects
Choose questions individually, practise online, or create a worksheet with matching answers. Counts include generated variants, not only authored originals.
Learn a mapping from examples with target labels
Supervised learning uses target-labelled training examples.
Define the task and error costs
Task definition determines evaluation and practical constraints.
A numeric delivery-time estimate
Regression predicts a numerical quantity.
An item category
Classification predicts a class or category.
72 %
72/100×100=72%.
It establishes a reference for improvement, cost and failure modes. A complex model is not useful merely because it is complex.
Credit measurable comparison and practical value.
Artificial intelligence is a broad field of methods for tasks involving perception, prediction, planning or language. Machine learning fits patterns from data rather than requiring every rule to be handwritten.
Grades 9–12Artificial intelligence is a broad field of methods for tasks involving perception, prediction, planning or language. Machine learning fits patterns from data rather than requiring every rule to be handwritten.
Grades 9–12An example is one unit of observation. Features are inputs available at prediction time; a label is the target used for supervised learning.
Grades 9–12An example is one unit of observation. Features are inputs available at prediction time; a label is the target used for supervised learning.
Grades 9–12The model fits its parameters on training data. Success there does not alone demonstrate generalization to new cases.
Grades 9–12The model fits its parameters on training data. Success there does not alone demonstrate generalization to new cases.
Grades 9–12Describe who uses the system, who may be affected, and who can correct mistakes. Automation does not remove human responsibility.
Grades 9–12Describe who uses the system, who may be affected, and who can correct mistakes. Automation does not remove human responsibility.
Grades 9–12A linear model predicts yhat=w·x+b. Linear refers to the parameters; feature transformations may make the relation nonlinear in original inputs.
UniversityA linear model predicts yhat=w·x+b. Linear refers to the parameters; feature transformations may make the relation nonlinear in original inputs.
UniversityGradient descent uses θnew=θ−η∇L(θ) with a positive step size η. The negative gradient is locally downhill for a differentiable objective.
UniversityGradient descent uses θnew=θ−η∇L(θ) with a positive step size η. The negative gradient is locally downhill for a differentiable objective.
UniversityA logistic model maps a real score through σ(z)=1/(1+exp(−z)). Interpreting a score as a reliable probability requires calibration assessment.
UniversityA logistic model maps a real score through σ(z)=1/(1+exp(−z)). Interpreting a score as a reliable probability requires calibration assessment.
UniversityA model can fit training-specific noise or patterns that do not generalize. Compare training and appropriately held-out performance.
UniversityA model can fit training-specific noise or patterns that do not generalize. Compare training and appropriately held-out performance.
UniversityA decision tree partitions feature space using a sequence of tests. Leaf predictions summarize the training examples reaching them.
UniversityA decision tree partitions feature space using a sequence of tests. Leaf predictions summarize the training examples reaching them.
UniversityClustering groups examples according to a representation and similarity rule. The groups are not automatically true natural categories.
UniversityClustering groups examples according to a representation and similarity rule. The groups are not automatically true natural categories.
UniversityPrincipal component analysis finds orthogonal directions of high variance in centred data under its standard formulation.
UniversityPrincipal component analysis finds orthogonal directions of high variance in centred data under its standard formulation.
UniversityA k-fold procedure rotates held-out folds to estimate performance for a specified training process. Splits and all fitted preprocessing must be controlled.
UniversityA k-fold procedure rotates held-out folds to estimate performance for a specified training process. Splits and all fitted preprocessing must be controlled.
UniversityA dense layer computes z=Wx+b followed by an activation. Dimensions must agree; each output has its own weighted combination.
UniversityA dense layer computes z=Wx+b followed by an activation. Dimensions must agree; each output has its own weighted combination.
UniversityA computational graph records how intermediate values compose. Forward evaluation computes values; reverse differentiation accumulates gradients.
UniversityA computational graph records how intermediate values compose. Forward evaluation computes values; reverse differentiation accumulates gradients.
UniversityScaled dot-product attention forms weights softmax(QKᵀ/sqrt(dk)) and uses them to combine value vectors. Dimensions and masks are part of the definition.
Advanced universityScaled dot-product attention forms weights softmax(QKᵀ/sqrt(dk)) and uses them to combine value vectors. Dimensions and masks are part of the definition.
Advanced universityAn embedding represents an item as a vector learned or designed for a task. Nearby vectors reflect the representation’s notion of similarity, not universal meaning.
UniversityAn embedding represents an item as a vector learned or designed for a task. Nearby vectors reflect the representation’s notion of similarity, not universal meaning.
UniversityAn autoregressive language model predicts successive tokens from context. A token may be a word piece, punctuation or another encoded unit.
Grades 9–12An autoregressive language model predicts successive tokens from context. A token may be a word piece, punctuation or another encoded unit.
Grades 9–12Retrieval-augmented generation first retrieves relevant source material, then conditions generation on selected context. It does not automatically make every answer correct.
UniversityRetrieval-augmented generation first retrieves relevant source material, then conditions generation on selected context. It does not automatically make every answer correct.
UniversityUse representative tasks with clear expected behaviour and deliberately include failure cases. Keep an evaluation set separate from prompt tuning.
UniversityUse representative tasks with clear expected behaviour and deliberately include failure cases. Keep an evaluation set separate from prompt tuning.
UniversityA useful AI application includes data handling, retrieval or model calls, validation, error handling and a user interface. The model is one component.
UniversityA useful AI application includes data handling, retrieval or model calls, validation, error handling and a user interface. The model is one component.
UniversityA model prediction is not a guarantee about an individual outcome.
UniversityA model prediction is not a guarantee about an individual outcome.
UniversityAverage across the three observations, not only the nonzero errors.
UniversityAverage across the three observations, not only the nonzero errors.
UniversityThe gradient and learning rate belong to the specified loss and parameterization.
UniversityThe gradient and learning rate belong to the specified loss and parameterization.
UniversityThe denominator counts predicted positives. This generator always gives a nonzero denominator.
Grades 9–12The denominator counts predicted positives. This generator always gives a nonzero denominator.
Grades 9–12The denominator counts actual positives, not predicted positives.
Grades 9–12The denominator counts actual positives, not predicted positives.
Grades 9–12This is a class-specific score, not a multiclass macro average or a cost-sensitive objective.
UniversityThis is a class-specific score, not a multiclass macro average or a cost-sensitive objective.
UniversityAssume one trainable bias per output and no parameter sharing.
UniversityAssume one trainable bias per output and no parameter sharing.
UniversityThis asks for a forward pass, not the derivative at the nondifferentiable point zero.
UniversityThis asks for a forward pass, not the derivative at the nondifferentiable point zero.
UniversitySoftmax normalizes scores; this alone does not prove probabilities are calibrated to real-world frequencies.
Advanced universitySoftmax normalizes scores; this alone does not prove probabilities are calibrated to real-world frequencies.
Advanced universityUse covariance eigenvalues sorted in descending order and a positive total variance.
UniversityUse covariance eigenvalues sorted in descending order and a positive total variance.
UniversityThe reference relevance labels and query set define the evaluation; recall does not measure answer correctness.
UniversityThe reference relevance labels and query set define the evaluation; recall does not measure answer correctness.
UniversityAssume strictly sequential stages and exclude retries or parallel overlap.
UniversityAssume strictly sequential stages and exclude retries or parallel overlap.
UniversityUnderstand learning tasks, datasets, honest evaluation, privacy and human responsibility. No paid AI tool is required.
Work through regression, gradients, classification and generalization using paper-based examples.
Compare trees, clustering, PCA and validation without confusing training performance with evidence of generalization.
Build from artificial neurons and backpropagation to attention and embedding geometry.
Practise clear prompting, retrieval-grounded workflows, evidence checking and permission-aware deployment.