Machine learning using the micro:bit

DASKEREL wrote this teaching and practice using Oak's sequence as a source. View the source lesson at Oak National Academy. Oak has not endorsed this material. Check Oak's terms and separate rights for third-party works.

Learning outcomes

Distinguish a programmed rule from a model trained on examples
Identify suitable low-stakes uses, error risks and human oversight

Vocabulary: artificial intelligence · machine learning · training example · prediction · human oversight

Retrieve previous learning

What does a sensor provide to a computer?

Check: An input measurement that software can interpret, subject to sensor and context limits.

Learn the concept

Artificial intelligence is a broad name for systems designed to perform tasks associated with human reasoning or perception. Machine learning is one approach: a model is fitted using examples so it can make predictions on new inputs. A fixed program such as 'if light level is below 40, show dim' follows an explicit threshold; a trained classifier may infer a boundary from labelled examples. Neither is automatically reliable or suitable for every decision.

For a classroom micro:bit project, consider classifying the device as still or gently moved from accelerometer readings. The training labels describe examples, not guaranteed truths for every future reading. Orientation, sensor differences and the way examples were collected can affect predictions. Use withheld test examples and inspect errors, including which class is missed.

Choose an application whose consequences are small and reversible, such as a demonstration indicator. Do not use a classroom movement model to judge a pupil's behaviour or make an attendance decision. Avoid collecting identities, offer a simulated-data route and a manual override, and say 'uncertain' or 'unavailable' when the evidence is insufficient. A teacher should approve the device use and inspect actual output.

Worked model

Compare an explicit light threshold with a classifier trained from labelled light readings.
Define a bounded use: show 'bright' or 'dim' on a demonstration display, with no personal record.
Identify tests for new lighting and sensor conditions, including an ambiguous reading.
Add a clear override and restrict the conclusion to the conditions actually tested.

Guided practice

A program says 'if reading < 40 then dim'. Was that threshold necessarily learned from examples?

Hint: Ask who specified the boundary.

Check the guided answer

No. As stated, a programmer wrote the boundary. A trained model would derive its decision rule from examples, although a person still chooses the data, labels and evaluation method.

Independent practice

Write your own response to every question before viewing the model answers. Your writing stays in this page session; this activity does not submit, store or grade it. Avoid including private information.

Teacher observed practical evidence

A teacher must apply the school's risk assessment, allergy policy and any individual healthcare plan before the practical. Observe and record the criteria during supervised work. Written answers cannot certify practical safety or skill.

Teacher observes collection and classification with a school-approved micro:bit, connected tool and accessible alternative where needed.
Learner retains consent-safe labelled data, a separate test set, model version and evidence of at least one correction.
Learner demonstrates the application and explains uncertain predictions and a human override; written accuracy claims are limited to the tested examples.

Find and correct a misconception

“A system called AI understands a person and can always be trusted to decide for them.”

Correction: A model predicts from limited data, can fail under changed conditions and needs purpose-specific evaluation and human oversight.

Extend and reflect

Compare a rule-based and example-trained approach to the same bright/dim task.

Guidance: Discuss what can be inspected, how boundaries are chosen, what tests each needs and what a missing reading should trigger.

Where this lesson could lead · Computer Science GCSE → Daskerel Computing and AI

This lesson practises: Distinguish a programmed rule from a model trained on examples. Algorithms, digital systems and safe practice support later software and technology work.

Next school subject: Computer Science GCSE. Check the subject and examination board that fit your school.
Later DASKEREL learning: Daskerel Computing and AI. Explore further study and build evidence from practical work.
Later credential exploration: AWS Certified Developer - Associate is an example in this later specialism. Check current awarding-body requirements before planning an exam.
Example career direction: Software Engineer. Explore organisations and local support.

This is an exploratory route, not a promise of admission, a qualification, certification eligibility or employment. Years 7–9 learning is an early foundation; any later credential may require substantial study, experience or local approval.

Year 9 · Algorithms and programming · original lesson pack

DASKEREL beta · internal AI review pending

Gather labelled sensor examples for a bounded model

DASKEREL wrote this teaching and practice using Oak's sequence as a source. View the source lesson at Oak National Academy. Oak has not endorsed this material. Check Oak's terms and separate rights for third-party works.

Learning outcomes

Define two observable sensor classes
Collect balanced, permission-safe examples with labels and conditions

Vocabulary: sensor · feature · label · class · sampling · consent

Retrieve previous learning

Why does a model's training accuracy not establish its performance on new examples?

Check: It may fit the training examples without generalising to independent inputs.

Learn the concept

A micro:bit can provide sensor readings such as acceleration or light level to a teacher-approved classroom tool. Define the task before collecting data: for example, classify a device held still versus gently moved. A label describes the intended class, and each recorded example should include the relevant sensor values and conditions.

Gather varied but safe examples for both classes. Ten still readings taken under one exact condition may not represent other orientations. Never instruct pupils to run, shake near others, expose the device to water or record private movement histories. A non-sensor or simulated data alternative should be available where physical handling is inaccessible.

Separate a planned test set before training. Check labels for mistakes and note how many examples each class has; a strongly unbalanced set can encourage a model to favour one label. The dataset is evidence for a limited task, not a general measure of a learner's behaviour.

Worked model

Define safe classes still and gently moved with teacher supervision.
Record several labelled readings per class across permitted orientations.
Inspect label errors and class counts without collecting identities.
Reserve independent examples for later testing and document the collection conditions.

Guided practice

A dataset has 20 'still' examples and 2 'moved' examples. What should be checked?

Hint: Consider whether the classes are represented comparably.

Check the guided answer

Check labels and collect more safe, varied moved examples if appropriate. Reserve independent examples from both classes; the imbalance could make a model favour still.

Independent practice

Write your own response to every question before viewing the model answers. Your writing stays in this page session; this activity does not submit, store or grade it. Avoid including private information.

Teacher observed practical evidence

A teacher must apply the school's risk assessment, allergy policy and any individual healthcare plan before the practical. Observe and record the criteria during supervised work. Written answers cannot certify practical safety or skill.

Teacher observes collection and classification with a school-approved micro:bit, connected tool and accessible alternative where needed.
Learner retains consent-safe labelled data, a separate test set, model version and evidence of at least one correction.
Learner demonstrates the application and explains uncertain predictions and a human override; written accuracy claims are limited to the tested examples.

Find and correct a misconception

“More sensor rows always make a model fairer.”

Correction: Repeated, mislabelled or one-sided rows may reinforce bias; coverage, labels and task relevance matter.

Extend and reflect

Plan a simulated-data route for a learner who cannot handle the device.

Guidance: Preserve the same labels and reasoning task while allowing accessible input and output.

Where this lesson could lead · Computer Science GCSE → Daskerel Computing and AI

This lesson practises: Define two observable sensor classes. Algorithms, digital systems and safe practice support later software and technology work.

Next school subject: Computer Science GCSE. Check the subject and examination board that fit your school.
Later DASKEREL learning: Daskerel Computing and AI. Explore further study and build evidence from practical work.
Later credential exploration: AWS Certified Developer - Associate is an example in this later specialism. Check current awarding-body requirements before planning an exam.
Example career direction: Software Engineer. Explore organisations and local support.

This is an exploratory route, not a promise of admission, a qualification, certification eligibility or employment. Years 7–9 learning is an early foundation; any later credential may require substantial study, experience or local approval.

Year 9 · Algorithms and programming · original lesson pack

DASKEREL beta · internal AI review pending

Train a model and test it on withheld examples

DASKEREL wrote this teaching and practice using Oak's sequence as a source. View the source lesson at Oak National Academy. Oak has not endorsed this material. Check Oak's terms and separate rights for third-party works.

Learning outcomes

Distinguish training and test examples
Calculate bounded test accuracy and inspect error types

Vocabulary: training · test set · prediction · accuracy · confusion

Retrieve previous learning

Why reserve test readings before training?

Check: They give a less biased check of how the model behaves on unseen examples.

Learn the concept

Use the labelled training examples to fit a two-class model in a teacher-approved tool. Do not include the reserved test examples during fitting or tuning. The tool's exact algorithm may be hidden; record its version and configuration rather than claiming to know an internal rule you have not inspected.

Run the withheld examples through the trained model. If it predicts 7 of 10 labels correctly, test accuracy on that ten-example set is 70%. Count which errors were still wrongly labelled moved and which were moved wrongly labelled still. A single percentage can hide a consequential imbalance.

A small test result is uncertain. Repeat under relevant conditions and compare model versions on the same untouched test set when possible. Do not reuse the test set for repeated tuning without acknowledging that it has become part of development. A human should review unexpected predictions before any action.

Worked model

Separate and label training and withheld examples.
Fit the model using only the training partition and record the tool/version.
Predict ten withheld labels, count seven correct and three errors for 70% on that set.
Inspect which classes were confused and state that this result is limited to ten examples.

Guided practice

A model gets 9/10 training examples and 6/10 withheld examples right. Which number is a better first check of new-input performance?

Hint: Choose examples not shown during fitting.

Check the guided answer

The withheld 6/10, or 60% on that small set, is the better first check. Training 9/10 shows fit to seen data and can overstate generalisation.

Independent practice

Write your own response to every question before viewing the model answers. Your writing stays in this page session; this activity does not submit, store or grade it. Avoid including private information.

Teacher observed practical evidence

A teacher must apply the school's risk assessment, allergy policy and any individual healthcare plan before the practical. Observe and record the criteria during supervised work. Written answers cannot certify practical safety or skill.

Teacher observes collection and classification with a school-approved micro:bit, connected tool and accessible alternative where needed.
Learner retains consent-safe labelled data, a separate test set, model version and evidence of at least one correction.
Learner demonstrates the application and explains uncertain predictions and a human override; written accuracy claims are limited to the tested examples.

Find and correct a misconception

“A model's 100% training score guarantees a safe device action.”

Correction: An independent relevant test and human oversight are needed; unseen conditions can still cause errors.

Extend and reflect

Compare two model versions with equal accuracy but different errors.

Guidance: Consider which class each misses and why the intended action changes the consequence.

Where this lesson could lead · Computer Science GCSE → Daskerel Computing and AI

This lesson practises: Distinguish training and test examples. Algorithms, digital systems and safe practice support later software and technology work.

Next school subject: Computer Science GCSE. Check the subject and examination board that fit your school.
Later DASKEREL learning: Daskerel Computing and AI. Explore further study and build evidence from practical work.
Later credential exploration: AWS Certified Developer - Associate is an example in this later specialism. Check current awarding-body requirements before planning an exam.
Example career direction: Software Engineer. Explore organisations and local support.

This is an exploratory route, not a promise of admission, a qualification, certification eligibility or employment. Years 7–9 learning is an early foundation; any later credential may require substantial study, experience or local approval.

Year 9 · Algorithms and programming · original lesson pack

DASKEREL beta · internal AI review pending

Improve a model with a documented hypothesis

DASKEREL wrote this teaching and practice using Oak's sequence as a source. View the source lesson at Oak National Academy. Oak has not endorsed this material. Check Oak's terms and separate rights for third-party works.

Learning outcomes

Identify a pattern in errors and a plausible intervention
Compare model versions on a fair test

Vocabulary: hypothesis · retraining · distribution shift · version · evaluation

Retrieve previous learning

What does 7/10 correct mean?

Check: The model was correct on seven of those ten tested examples, or 70% for that set; it does not establish universal accuracy.

Learn the concept

Improvement starts with a diagnosed problem. If a still/moved classifier fails when the device is tilted, the hypothesis may be that training examples covered only one orientation. Add safe, labelled examples across permitted tilts rather than randomly changing every setting.

Record a baseline model, the planned change and the new model version. Keep a genuinely untouched evaluation set or collect new independent examples after repeated tuning. Compare class-specific errors as well as overall correct count. A one-example improvement may be noise, especially with a very small test set.

Do not solve an error by collecting unnecessary sensitive data. Simulated or anonymous examples may suffice. If a model remains unreliable for its intended use, the responsible choice can be to reduce its role or use a manual rule. 'More AI' is not itself a success criterion.

Worked model

Record baseline failures on tilted-device examples.
Write a hypothesis about missing orientation variety and add safe labelled training readings.
Train a new version while keeping an independent evaluation set.
Compare error counts by class and state whether evidence supports the improvement.

Guided practice

A version rises from 7/10 to 8/10 on a tiny set. Can it be declared reliable?

Hint: Think about uncertainty and reuse of test examples.

Check the guided answer

No. It improved by one example on that set, but the sample is small and may have influenced tuning. Test more relevant, independent examples and inspect which errors remain.

Independent practice

Write your own response to every question before viewing the model answers. Your writing stays in this page session; this activity does not submit, store or grade it. Avoid including private information.

Teacher observed practical evidence

A teacher must apply the school's risk assessment, allergy policy and any individual healthcare plan before the practical. Observe and record the criteria during supervised work. Written answers cannot certify practical safety or skill.

Teacher observes collection and classification with a school-approved micro:bit, connected tool and accessible alternative where needed.
Learner retains consent-safe labelled data, a separate test set, model version and evidence of at least one correction.
Learner demonstrates the application and explains uncertain predictions and a human override; written accuracy claims are limited to the tested examples.

Find and correct a misconception

“The newest model version is automatically the best.”

Correction: Compare versions against relevant independent evidence and consequences, not version number.

Extend and reflect

Design a model-version log entry.

Guidance: Record data change, model configuration, test set, class-specific errors, date and decision.

Where this lesson could lead · Computer Science GCSE → Daskerel Computing and AI

This lesson practises: Identify a pattern in errors and a plausible intervention. Algorithms, digital systems and safe practice support later software and technology work.

Next school subject: Computer Science GCSE. Check the subject and examination board that fit your school.
Later DASKEREL learning: Daskerel Computing and AI. Explore further study and build evidence from practical work.
Later credential exploration: AWS Certified Developer - Associate is an example in this later specialism. Check current awarding-body requirements before planning an exam.
Example career direction: Software Engineer. Explore organisations and local support.

This is an exploratory route, not a promise of admission, a qualification, certification eligibility or employment. Years 7–9 learning is an early foundation; any later credential may require substantial study, experience or local approval.

Year 9 · Algorithms and programming · original lesson pack

DASKEREL beta · internal AI review pending

Connect a prediction to a safe application action

DASKEREL wrote this teaching and practice using Oak's sequence as a source. View the source lesson at Oak National Academy. Oak has not endorsed this material. Check Oak's terms and separate rights for third-party works.

Learning outcomes

Map sensor input through model prediction to output
Provide an uncertainty and override route

Vocabulary: pipeline · input · prediction · threshold · override

Retrieve previous learning

Why might a simpler rule beat a model?

Check: A stable threshold can be easier to explain, test and maintain for a narrowly defined task.

Learn the concept

A data-driven application has a chain: sensor reading → preprocessing → model prediction → decision rule → output. For a fictional desk device that suggests a stretch break when it detects stillness, the prediction is a suggestion, not a medical judgement or a command to move. The learner should be able to see and dismiss it.

Define how the app handles missing or uncertain input. A sensor failure should display 'reading unavailable' rather than pretending the device is still. If a model's confidence value is provided, it is not automatically a calibrated probability; test what it means in the tool. Avoid collecting histories of named pupils.

Test a normal reading, an ambiguous movement and a disconnected sensor. Check the actual output on the device or accessible alternative. A human override and a clear explanation allow a user to correct the system. A teacher must supervise physical setup and confirm the program behaves as documented.

Worked model

Write the fictional app's input, prediction and non-medical suggestion.
Specify a clear dismiss or override action and a missing-sensor message.
Run teacher-supervised normal, ambiguous and disconnected tests.
Record outputs, defects and corrections without personal movement histories.

Guided practice

The sensor disconnects, but the display says 'You have been still for an hour'. What is wrong?

Hint: No reading was available to support the claim.

Check the guided answer

The app is inventing a state from missing data. It should say the reading is unavailable and avoid the suggestion until it has valid input or a safe fallback.

Independent practice

Write your own response to every question before viewing the model answers. Your writing stays in this page session; this activity does not submit, store or grade it. Avoid including private information.

Teacher observed practical evidence

A teacher must apply the school's risk assessment, allergy policy and any individual healthcare plan before the practical. Observe and record the criteria during supervised work. Written answers cannot certify practical safety or skill.

Teacher observes collection and classification with a school-approved micro:bit, connected tool and accessible alternative where needed.
Learner retains consent-safe labelled data, a separate test set, model version and evidence of at least one correction.
Learner demonstrates the application and explains uncertain predictions and a human override; written accuracy claims are limited to the tested examples.

Find and correct a misconception

“A prediction is a fact about the person using the device.”

Correction: It is a model output with uncertainty and limited context; application design must account for error and user agency.

Extend and reflect

Sketch a text-only interface for the same light task.

Guidance: Preserve meaning and override without relying on flashing LEDs or colour alone.

Where this lesson could lead · Computer Science GCSE → Daskerel Computing and AI

This lesson practises: Map sensor input through model prediction to output. Algorithms, digital systems and safe practice support later software and technology work.

Next school subject: Computer Science GCSE. Check the subject and examination board that fit your school.
Later DASKEREL learning: Daskerel Computing and AI. Explore further study and build evidence from practical work.
Later credential exploration: AWS Certified Developer - Associate is an example in this later specialism. Check current awarding-body requirements before planning an exam.
Example career direction: Software Engineer. Explore organisations and local support.

This is an exploratory route, not a promise of admission, a qualification, certification eligibility or employment. Years 7–9 learning is an early foundation; any later credential may require substantial study, experience or local approval.

Year 9 · Algorithms and programming · original lesson pack

DASKEREL beta · internal AI review pending

Document a model's evidence, limits and decision

DASKEREL wrote this teaching and practice using Oak's sequence as a source. View the source lesson at Oak National Academy. Oak has not endorsed this material. Check Oak's terms and separate rights for third-party works.

Learning outcomes

Report data, tests and errors honestly
Recommend use, revision or rejection for a bounded task

Vocabulary: model card · provenance · test accuracy · failure mode · decision

Retrieve previous learning

What does a missing sensor reading require?

Check: An explicit unavailable or safe fallback state, not an invented prediction.

Learn the concept

A model record should name its intended task, data source and consent basis, labels, training and independent test method, tool version and observed errors. Give counts as well as percentages. '8/10 correct on ten classroom examples' communicates more than '80% accurate' with no context.

State known failure modes such as device tilt, lighting or a class underrepresented in training. Explain whether the app uses a human override, how missing input is handled and whether the data needs deletion after the project. A model may be unsuitable for a task even if one test score looks strong.

Finish with a decision supported by evidence: use only for a low-stakes demonstration, collect additional permission-safe tests, revise the design, or reject deployment. This is a classroom evaluation, not certification that a model is fair or safe in wider use. A teacher should inspect the saved data, model and test record.

Worked model

Record the still/moved model's bounded purpose and non-personal collection method.
Report a fictional 8/10 withheld result with both errors described.
State tilt as a failure condition and show the manual override.
Recommend demonstration-only use pending larger independent tests, then save the model record.

Guided practice

A report says only 'excellent accuracy'. Name two concrete missing pieces.

Hint: A reader needs to inspect the test and error pattern.

Check the guided answer

Give the correct/total count on an independent set and describe which classes were misclassified. The phrase alone hides both sample size and failure type.

Independent practice

Write your own response to every question before viewing the model answers. Your writing stays in this page session; this activity does not submit, store or grade it. Avoid including private information.

Teacher observed practical evidence

A teacher must apply the school's risk assessment, allergy policy and any individual healthcare plan before the practical. Observe and record the criteria during supervised work. Written answers cannot certify practical safety or skill.

Teacher observes collection and classification with a school-approved micro:bit, connected tool and accessible alternative where needed.
Learner retains consent-safe labelled data, a separate test set, model version and evidence of at least one correction.
Learner demonstrates the application and explains uncertain predictions and a human override; written accuracy claims are limited to the tested examples.

Find and correct a misconception

“A model card is proof that an ML system is approved for any use.”

Correction: It records evidence and limits; a responsible decision still depends on the intended context and review.

Extend and reflect

Compare a low-stakes classroom demonstration with a high-stakes decision.

Guidance: Explain why evidence, oversight, rights and harm controls must rise with consequences.