IB COMPUTER SCIENCE | A4 ASSESSMENT
MACHINE LEARNING: END OF TOPIC REVIEW
Check learning fundamentals, hardware, labelled HL methods and ethical implications through scored retrieval, applied tasks and developed written responses.
BEFORE YOU BEGIN
Use two or more sessions if needed: first the automatic checks and applied tasks, then the written responses. Allow time to show working and develop discussion answers. Common SL/HL content is A4.1 and A4.4; A4.2 and A4.3 are HL-only.
Sections A and B award one point for each fully correct automatic item, with no partial credit within an item. Section C has separately suggested content marks. These are original practice questions, not an official IB paper, grade boundary or prediction.
Choose your level below. Write calculations, diagrams and Section C answers in your exercise book before revealing guidance. Answers and selections last while the page remains open.
Automatic total: 48 points. Suggested written total: 84 marks, assessed separately.
C | WRITTEN COMMUNICATION AND SUBJECT KNOWLEDGE
Write in your exercise book before revealing the guidance. The table gives practical guidance based on the IB command terms supplied in the guide; it is not the Cambridge command-word table.
In particular, IB compare focuses on similarities, while distinguish asks for differences. If you need to address both, the question can request each separately. All main commands below are highlighted.
| IB command term | What to provide in this review | Writing approach |
|---|---|---|
| Calculate | A numerical result with relevant working stages | Show the calculation and units |
| Describe | A detailed account | Give features or an ordered process, not only a name |
| Explain | A detailed account including reasons or causes | Link ideas with because, therefore, resulting in or leading to |
| Compare | Similarities, referring to both items throughout | Both…; each…; in both approaches… |
| Compare and contrast | Similarities and differences, referring to both throughout | Use both for similarities and whereas for differences |
| Discuss | A considered, balanced review of relevant arguments or factors | Develop strengths and limits in context and reach a supported judgement |
| Distinguish | Clear differences between the concepts | Whereas…; unlike…; one uses distances whereas the other follows learned tests… |
| Construct | Information displayed in diagrammatic or logical form | Supply the requested table, map or connected diagram |
| Deduce | A conclusion reached from the supplied information | Use the map or data as evidence |
| Justify | Valid reasons or evidence for a choice | Connect the choice to the stated requirement |
| Evaluate | An appraisal weighing strengths and limitations | Balance relevant factors and reach a supported judgement |
A connective earns no automatic mark by itself: it must express an accurate relationship. For example, “a large unscaled feature can dominate distance, therefore K-NN may select misleading neighbours under that representation” explains a mechanism and consequence. Suggested marks reward subject content, not extra unofficial grammar marks.
C1 | LEARNING SETTINGS
6 suggested marks | COMMON SL/HL
Supervised and unsupervised approaches both use recorded feature examples to model patterns. Supervised training uses known targets, such as journey durations; unsupervised clustering finds groups without known target-group labels guiding the fit. Reinforcement learning instead improves decisions through interaction: a robot chooses an action, receives a reward and next observation, and uses experience to improve its policy toward expected cumulative return. A supplied correct action for every situation is not required.
SUGGESTED MARKING POINTS
- Gives one valid similarity between supervised and unsupervised learning.
- Explains use of known targets in supervised fitting.
- Explains unsupervised structure discovery without supplied target-group labels.
- Identifies agent actions and an interacting environment in the robot example.
- Describes reward and next-observation feedback.
- Distinguishes learning a decision policy toward cumulative return from the other settings.
Writing focus: Refer to both approaches for similarities and differences; use whereas for a genuine contrast.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC2 | JUSTIFY HARDWARE
6 suggested marks | COMMON SL/HL
The robot needs compatible local processing and sensor I/O to meet a response target without relying on a remote round trip. Its execution memory must hold the model and working values, while persistent storage retains code and model files. Power and heat constrain the device. The remote training job may use suitable parallel accelerators with host processing, substantial working memory and storage for data/checkpoints. A balanced input pipeline and interconnects matter because communication or loading can limit scaling; more devices do not guarantee proportional speed.
SUGGESTED MARKING POINTS
- Connects local compatible processing or sensor I/O to the robot’s response requirement.
- Distinguishes model/runtime execution memory from persistent storage.
- Justifies power, heat or local communication constraints.
- Justifies supported training compute or parallel accelerators for the large workload.
- Includes training data/checkpoint storage or adequate working memory.
- Explains a valid scaling bottleneck such as input supply or synchronisation.
Writing focus: Link a resource to its purpose with because or therefore; do not offer an unexplained shopping list.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC3 | SCHOOL SUPPORT ETHICS
6 suggested marks | COMMON SL/HL
A model could help teachers identify needs earlier and manage limited support capacity. However, past referrals may reflect unequal access, so reproducing them can overlook underrepresented students. Combining personal records also creates privacy and inference concerns. Necessary data, group-error investigation and accessible correction can help. Reviewers need time, context and override authority rather than a requirement to accept the score. I would support a limited decision-support trial only if it improves access compared with the existing process and provides effective remedies; aggregate accuracy alone does not justify it.
SUGGESTED MARKING POINTS
- Develops an intended benefit in the support context.
- Explains how historical labels or coverage can cause unequal treatment.
- Develops a privacy or inference concern.
- Proposes a relevant response and explains how it helps.
- Explains a limitation of review or safeguards.
- Reaches a supported conditional judgement in context.
Writing focus: Balance benefits and concerns, then state conditions rather than ending with “there are pros and cons”.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC4 | REASSESS AN EXPANDED TECHNOLOGY
6 suggested marks | COMMON SL/HL
Translation may support independence, but recognition and profiling introduce a different purpose and affect bystanders who did not choose the service. Inferences can be wrong or discriminatory and can expose identity or private circumstances. The wearer’s agreement is not sufficient evidence about everyone affected. Consider restricting unnecessary capture, practical alternatives and whether identification is justified at all. Because capability and consequences have changed, a named review process with affected people and authority to restrict the feature is needed; the original judgement should not automatically extend to this use.
SUGGESTED MARKING POINTS
- Explains a contextual translation benefit.
- Identifies the changed purpose or capability.
- Explains a bystander privacy or rights consequence.
- Explains an equity, error or discrimination concern.
- Considers an alternative or safeguard with a limit.
- Justifies a new review and responsibility rather than automatic inherited approval.
Writing focus: Make the change explicit, then use because to connect it to the need for reassessment.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC5 | ONLINE COMMUNICATION
6 suggested marks | COMMON SL/HL
Fast moderation can reduce exposure to abuse while anonymous participation can protect vulnerable speakers. However, context or dialect errors can suppress legitimate expression, and missed abuse still harms users. Requiring public names may create privacy or exclusion risks without eliminating harassment. Clear proportionate rules, context-sensitive review and protected reporting or appeals can help, but workload and delays limit effectiveness. A defensible policy needs evidence on both missed harm and false restriction, meaningful user remedies and a reasoned balance for the platform.
SUGGESTED MARKING POINTS
- Develops the safety benefit of timely moderation.
- Explains a legitimate benefit of anonymity or pseudonymity.
- Explains a false-removal or missed-abuse consequence.
- Explains a privacy or exclusion cost of universal identification.
- Assesses a relevant response including its practical limit.
- Reaches a supported contextual balance rather than an absolute claim.
Writing focus: Discuss competing harms without assuming either total identification or no moderation is a complete solution.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC6 | FUTURE CAPABILITY AND DAILY LIFE
6 suggested marks | COMMON SL/HL
Some long-lived confidential data could be exposed if a sufficiently capable future quantum computer defeats its relevant public-key protection. This is conditional, not a claim that all encryption is broken today, so review data lifetimes and qualified transition options proportionately. A compulsory VR activity can broaden learning but excludes people if equipment, accessibility or equivalent routes are inadequate; detailed behavioural capture also needs justification. Both cases show why changing capability, necessity and alternatives need review before harm, with accountable owners and evidence rather than a blanket rejection or automatic approval of innovation.
SUGGESTED MARKING POINTS
- Explains a conditional long-lived-data quantum concern.
- Avoids claiming all current encryption is broken or a universal arrival date.
- Develops a VR learning benefit.
- Explains an access, accessibility or behavioural-privacy burden.
- Assesses a proportionate response or alternative.
- Connects both cases to accountable continuing reassessment.
Writing focus: Separate future possibility from established evidence and make the conditional conclusion explicit.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC7 | HL: PREPARE THE DATA
6 suggested marks | HL ONLY
Remove verified repeat imports of the same event, not merely rows sharing values. Confirm source units and convert consistently. Investigate missingness before deleting rows; a training-derived imputer estimates absent distances but does not recover ground truth. Retain the verified long journey if it belongs to the intended population. Select or reduce inputs only where appropriate to preserve useful information. Fit imputation, scaling and selection within training folds, because using held-out statistics or targets leaks information and can make evaluation optimistic.
SUGGESTED MARKING POINTS
- Explains event-based duplicate handling.
- Explains verified unit correction.
- Justifies a missing-value response with a limitation.
- Distinguishes a valid relevant outlier from an error.
- Explains preservation of useful inputs during selection or reduction.
- Explains training-only fitting as leakage prevention.
Writing focus: Connect each action to the specific problem rather than listing cleaning method names.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC8 | HL: REGRESSION AND FIT
6 suggested marks | HL ONLY
At 4.5 km, prediction=2.4×4.5+3.8=14.6 minutes. R²=1−4.4/62≈0.929 on the supplied training observations. Slope means an extra 2.4 predicted minutes per kilometre; intercept is the model prediction at zero distance, not necessarily a measured fixed delay. R² compares squared error with variation around the mean, not percentage exactly correct. Seven kilometres is extrapolation, so the relationship may change and training fit does not establish validity there.
SUGGESTED MARKING POINTS
- Correct substitution and prediction 14.6.
- Correct R² calculation approximately 0.929.
- Explains slope as predicted change per input unit.
- Explains intercept at zero with an appropriate limit.
- Interprets R² as relative squared-error fit, not exact-match accuracy.
- Explains extrapolation and training-evaluation limitation.
Writing focus: Show working, units and what the score describes, rather than just two final numbers.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC9 | HL: TWO CLASSIFICATION MECHANISMS
6 suggested marks | HL ONLY
Both use labelled examples to predict predefined categories and need appropriate independent evaluation. K-NN retains examples, computes input distances, selects k neighbours and combines their known labels, making scale and meaningful similarity important. A tree learns feature tests and follows one matching path to a leaf, so its output need not require neighbour-distance calculations. A small k can be noisy or broad k can hide small groups; a deep tree can fit sample-specific detail. Neither has a universally best configuration.
SUGGESTED MARKING POINTS
- States a valid shared use of labelled data and categorical outputs.
- Describes K-NN distance and neighbour selection.
- Describes combining selected labels.
- Describes a tree’s learned tests and one root-to-leaf path.
- Explains a K-NN representation or configuration limitation.
- Explains a tree complexity limitation or explicit mechanism contrast.
Writing focus: Refer to both throughout; use both for similarities and whereas for differences.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC10 | HL: METRICS AND SELECTION
6 suggested marks | HL ONLY
Precision=16/(16+8)=66.7%; recall=16/(16+4)=80%. Their denominators answer different questions. A missed defect and unnecessary inspection can have different costs, so a headline metric may not represent the objective. A candidate exceeding memory or response requirements can be ineligible despite a good score. Use development validation for family and configuration choices, then independent final testing of the chosen pipeline. Equal scores or a small gap do not prove equivalent errors or stable superiority.
SUGGESTED MARKING POINTS
- Correct precision with predicted-positive denominator.
- Correct recall with actual-positive denominator.
- Explains a contextual FP/FN cost difference.
- Explains hard eligibility constraints.
- Separates validation choices from final independent testing.
- Explains a relevant score or uncertainty limitation.
Writing focus: Use the named positive class and connect metric limitations to the selection decision.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC11 | HL: GROUPS AND ASSOCIATIONS
6 suggested marks | HL ONLY
Clustering groups customers or observations by feature similarity without target-group labels. Association rules relate co-occurring items or attribute-value facts within defined transactions. Support=12/100=12%; confidence=12/20=60%; lift=0.6/0.3=2. The consequent occurs twice as frequently among antecedent transactions as overall in these data. This does not establish causation, time order or a permanent customer identity; counts and independent stability require scrutiny.
SUGGESTED MARKING POINTS
- Clearly distinguishes grouped observations from co-occurring attributes/items.
- Explains the observation/transaction context or absence of supplied group labels.
- Correct support 12%.
- Correct confidence 60%.
- Correct lift 2 with baseline comparison.
- States a valid interpretation limit.
Writing focus: Keep the three denominators clear and do not turn the rule arrow into cause or sequence.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC12 | HL: REWARD LEARNING AND GENETIC SEARCH
6 suggested marks | HL ONLY
The longer route has raw return −1+8=7 compared with 2, so exploration can reveal a useful delayed outcome. Exploitation selects an action currently believed best, which can preserve incomplete early knowledge if alternatives are never tried. A genetic optimiser instead evaluates complete route chromosomes, selects parents, combines compatible order information and mutates some candidates. Repeated population evaluation and replacement continue to a stopping rule. This is not an RL state transition, and neither a short training run nor a stopped genetic search proves a global optimum.
SUGGESTED MARKING POINTS
- Calculates or correctly explains the delayed raw return 7 versus 2.
- Explains exploration as information gathering.
- Explains exploitation and its dependence on current estimates.
- Describes a population of complete route candidates and fitness evaluation.
- Describes selection plus representation-compatible crossover/mutation.
- Includes repeated replacement/termination and a valid guarantee limit.
Writing focus: Distinguish an RL step from a GA generation while connecting both to their actual objective.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC13 | HL: ANN STRUCTURE
6 suggested marks | HL ONLY
The perceptron sketch should show inputs, weighted connections, an added bias, an activation and output. The MLP has two input values, two hidden computed units and one output unit, with connections from each layer to the next. Nonlinear hidden transformations allow combinations beyond one affine boundary; stacking only affine mappings remains affine. Forward propagation computes outputs with current parameters; training uses loss and a learning procedure to adjust weights and biases. The architecture alone does not establish performance.
SUGGESTED MARKING POINTS
- Labels perceptron inputs and weights.
- Labels bias, activation and output in the correct computation order.
- Shows the 2–2–1 input/hidden/output structure.
- Shows appropriate inter-layer connections.
- Explains nonlinearity beyond an affine mapping.
- Distinguishes forward calculation from parameter training.
Writing focus: Make the diagrams readable; text should explain the relationship rather than replace missing labels.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentC14 | HL: CNN HIERARCHIES
6 suggested marks | HL ONLY
An image is represented as a spatial tensor. Shared local kernels produce feature maps, activation introduces nonlinearity and pooling or strided operations can reduce spatial detail. Later layers combine earlier responses over broader receptive fields; a suitable head produces class scores. Kernel size changes the local window, while stride changes sampling positions and output resolution. A loss compares outputs with targets, and training updates filters and later parameters. The hierarchy is learned and task-dependent, not a guarantee that every unit has a named human concept or every prediction is reliable.
SUGGESTED MARKING POINTS
- Describes spatial input and local shared filter computation.
- Explains nonlinear activation.
- Describes pooling/downsampling and a suitable classification head.
- Explains combined representations or broader receptive fields.
- Distinguishes effects of kernel size and stride.
- Links loss and parameter updating to adaptive learning.
Writing focus: Connect stages and distinguish architecture settings from learned kernel values.
Equivalent accurate wording and methods can be valid. Use the points as self-assessment guidance, not an official IB markscheme.
Revisit related contentCHECK THE QUALITY OF YOUR WRITING
Did you answer the command? For explain, check the reason or cause; for compare, address both items through similarities; for compare and contrast, include differences too; for discuss, balance relevant factors in context; for distinguish, make the differences explicit; for evaluate, weigh strengths and limitations and support your conclusion.
For construct, check labels, connections and completeness. For calculate, show enough working to follow your method and label units. Select two weak responses and rewrite them after revising. Keep the written total separate from your automatic result.
YOUR NEXT REVISION STEP
Submit the automatic sections for targeted revision links.
Record your own written marks using the suggested points. Choose two weaknesses and attempt new examples after revising. Use the A4 revision page for worked content and a personal flashcard list, or return to the IB course index.