Machine learning fundamentals: types and hardware – IB Computer Science exam questions with mark scheme

1Core3 marks

Describe the difference between supervised, unsupervised and reinforcement learning. [3]

Supervised: trained on labelled data (inputs with known correct outputs)(1)
Unsupervised: finds patterns/clusters in unlabelled data(1)
Reinforcement: an agent learns by trial and error from rewards/penalties for its actions(1)
Answer: Labelled data; unlabelled patterns; reward-based trial and error.
Examiner tip: Give the type of data (labelled/unlabelled) or feedback (reward) that defines each type.

2Core4 marks

Identify the type of machine learning being used in each of the following scenarios.

(a) An email provider trains a model on two million emails, each already tagged as “spam” or “not spam”, so that it can sort new emails. [1]

(b) A supermarket groups its customers according to their buying habits. No groups are defined in advance. [1]

(c) A warehouse robot learns to pick up parcels. It receives +1 point each time it lifts a parcel successfully and −1 point each time it drops one. [1]

(d) A developer takes a model that has already been trained to recognise everyday objects and retrains only its final layers on 500 photographs of diseased leaves, so that it can identify plant diseases. [1]

(a) Supervised (learning) (accept classification)(1)
(b) Unsupervised (learning) (accept clustering)(1)
(c) Reinforcement (learning)(1)
(d) Transfer (learning) (do not accept deep learning on its own)(1)
Answer: (a) Supervised (b) Unsupervised (c) Reinforcement (d) Transfer learning
Examiner tip: In (d) the clue is a model already trained on one task being reused for another; “deep learning” does not score.

3Core3 marks

A small start-up wants to build an app that recognises different breeds of dog from photographs, but it only has 3000 labelled photographs and a limited budget.

(a) Describe what is meant by transfer learning. [2]

(b) Outline one reason why transfer learning is suitable for this start-up. [1]

(a) A model that has already been trained on one (large) task/dataset (a pre-trained model) …(1)
(a) … is reused/adapted as the starting point for a different but related task, e.g. by fine-tuning/retraining only some (final) layers on the new data(1)
(b) It needs far less (labelled) training data, so 3000 photos is enough / it needs much less training time or computing power, so costs are lower (accept: the pre-trained model already recognises general features such as edges and shapes) (do not accept “it is faster” without a reason linked to data, training time or cost)(1)
Answer: (a) Reusing a model pre-trained on one task as the starting point for a related task, fine-tuning it on new data. (b) Needs less data / less training time and computing cost.
Examiner tip: Say what is reused (a pre-trained model) and what it is reused for (a different, related task); “learning from another model” is too vague.

Stuck on Machine learning fundamentals: types and hardware?

Get it sorted in a free 30-minute session

We work through your weakest questions on this topic on Zoom. I reply the same day.

Free 30-min session on this topic Send me 5 questions a day

4Standard3 marks

A research team trains a deep neural network. Training involves multiplying very large matrices of numbers millions of times.

Explain why the team would train the network on a graphics processing unit (GPU) rather than on a central processing unit (CPU). [3]

A GPU has many (hundreds/thousands of) simpler cores, whereas a CPU has a small number of more complex cores(1)
so a GPU can carry out many calculations at the same time / in parallel (SIMD)(1)
training is mainly the same (matrix/multiply–add) operation repeated on many independent values, which can be split across the cores, so training is completed much faster (accept GPUs have high memory bandwidth to feed large amounts of data to the cores)(1)
Answer: GPUs have thousands of cores that perform the same matrix operations in parallel, so training finishes much faster than on a CPU's few cores.
Examiner tip: “A GPU is more powerful” scores nothing; the mark comes from many cores working in parallel on repeated, independent calculations.

5Hard4 marks

A company designs a smart security camera that runs a trained image-recognition model. It is considering two types of chip to run the model: a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC).

Compare the use of an FPGA and an ASIC for this purpose. [4]

An FPGA can be reprogrammed/reconfigured after manufacture, so the model/circuit can be updated later, whereas an ASIC’s circuit is fixed once it is made(1)
An ASIC is custom-built for one task, so it gives the highest performance / lowest power use per operation (better than an FPGA)(1)
An ASIC has very high design/development (set-up) cost and long development time, so it is only worthwhile when very many units are made; its cost per unit is then low(1)
An FPGA has lower development cost/time, so it suits smaller production runs or prototypes, but each chip is more expensive/less efficient than an ASIC in mass production (a TPU is an example of an ML ASIC; naming it alone earns no mark)(1)
Answer: FPGA: reprogrammable, cheaper/faster to develop, good for prototypes or changing models. ASIC: fixed, fastest and most energy-efficient, cheap per unit only in large volumes.
Examiner tip: “Compare” needs both chips in each point; one-sided statements about only FPGA or only ASIC lose marks.

6Standard6 marks

A student wants to train an image-classification model. The training requires a total of $5.4 \times 10^{17}$ floating-point operations. The table shows the peak processing speed of three options. 1 TFLOPS = $10^{12}$ floating-point operations per second.

Option	Peak speed (TFLOPS)	Cost
Student’s laptop CPU	0.4	already owned
Desktop GPU	30	already owned
Cloud TPU instance	250	£8.00 per hour

(a) Calculate the time, in hours, to train the model on the laptop CPU. [2]

(b) Calculate the time, in hours, to train the model on the desktop GPU. [1]

(d) Suggest one reason why the real training time is likely to be longer than the values you calculated. [1]

(a) $5.4\times10^{17} \div (0.4\times10^{12})$ = $1.35\times10^{6}$ s (seen or implied)(1)
(a) 375 (hours) (dividing by 3600)(1)
(b) 5 (hours)(1)
(c) time = $5.4\times10^{17}\div(250\times10^{12})$ = 2160 s = 0.6 hours (allow 36 minutes)(1)
(c) £4.80 (follow through: their hours × 8; allow £8 if the student rounds up to a whole billed hour and says so)(1)
(d) e.g. the processor cannot run at full/peak speed all the time / time is spent moving data between memory/storage and the processor / uploading the data set to the cloud / other processes share the hardware / overheads such as loading data and saving checkpoints (any one reasonable)(1)
Answer: (a) 375 hours (b) 5 hours (c) 0.6 h × £8 = £4.80 (d) e.g. data transfer/memory bottlenecks mean the hardware does not run at full speed.
Examiner tip: Convert TFLOPS to operations per second (×10¹²) and seconds to hours (÷3600) before giving the answer.

7Hard8 marks

A smart doorbell uses a machine-learning model to recognise the faces of visitors. The doorbell camera captures 15 frames per second and each frame is 200 KB.

The manufacturer is choosing between two designs:

Cloud design: every frame is uploaded to a cloud server, which runs the model and sends back the result.
Edge design: a small low-power AI chip inside the doorbell runs the model, and only an alert is sent when an unknown face is detected.

(a) Calculate the amount of data, in GB, that the cloud design would upload in one hour. Use 1 GB = $10^{6}$ KB. [2]

(b) Explain two advantages of the edge design for this doorbell. [4]

(a) 15 × 200 × 3600 = 10 800 000 KB (seen or implied)(1)
(a) 10.8 (GB)(1)
(b) Lower latency / faster response …(1)
(b) … because frames do not travel to a remote server and back, so the owner/door can react in real time to a visitor(1)
(b) Better privacy/security … / far less bandwidth used … / works without an internet connection …(1)
(b) … because images of faces stay on the device and are not sent over the internet / only small alerts are sent instead of about 10.8 GB per hour / it can still recognise faces if the Wi-Fi or cloud service goes down (expansion must match the second advantage)(1)
(c) The edge chip has limited processing power/memory/battery (power budget) … (accept: model is harder to update or retrain on the device)(1)
(c) … so a smaller/simplified (compressed) model must be used, which may be less accurate (at recognising faces) / the device may become hot or drain its battery (accept: each doorbell must be updated individually)(1)
Answer: (a) 10.8 GB (b) e.g. lower latency for a real-time response; privacy because face images stay on the device; less bandwidth; works offline. (c) Limited power/memory means a smaller, less accurate model.
Examiner tip: Each “explain” advantage needs a reason linked to the doorbell, not just a word such as “faster” or “secure”.

8Standard5 marks

A car manufacturer is developing a self-driving car. It uses deep learning to recognise pedestrians in camera images, and reinforcement learning in a driving simulator to learn when to change lanes.

(a) Outline what is meant by deep learning. [2]

(b) Explain why reinforcement learning is suitable for learning when to change lanes. [2]

(a) Machine learning that uses (artificial) neural networks with many (hidden) layers(1)
(a) Each layer learns/extracts increasingly complex features automatically from (large amounts of) raw data, e.g. edges → shapes → a person (accept: needs very large data sets and high computing power)(1)
(b) There is no labelled “correct” answer for every situation; instead the agent tries actions/learns by trial and error in the simulator(1)
(b) It receives rewards (e.g. for safe, smooth progress) or penalties (e.g. for collisions) and learns a strategy/policy that maximises total reward; the simulator lets mistakes happen safely(1)
(c) e.g. game-playing AI (chess, Go, video games) / robot movement or control / traffic-light timing / energy management in data centres / trading strategies (any one reasonable; do not accept self-driving cars)(1)
Answer: (a) ML using neural networks with many layers that learn features automatically. (b) Learns by trial and error from rewards/penalties, with no labelled answers needed. (c) e.g. game-playing AI or robot control.
Examiner tip: For deep learning, “many layers” is the key idea; just saying “learns deeply” repeats the question.

9Hard5 marks

A research group trains a large language model on a high-performance computing (HPC) cluster. The graph shows how the training time changes as more GPUs in the cluster are used.

(a) State the training time when 4 GPUs are used. [1]

(b) The speed-up is the training time on 1 GPU divided by the training time on $n$ GPUs. Calculate the speed-up when 16 GPUs are used. [1]

(c) The efficiency is the speed-up divided by the number of GPUs. Calculate the efficiency, as a percentage, when 16 GPUs are used. [1]

(d) Explain why doubling the number of GPUs does not halve the training time. [2]

(a) 27 hours(1)
(b) 96 ÷ 9.6 = 10(1)
(c) 10 ÷ 16 × 100 = 62.5% (follow through from (b); accept 62% or 63%)(1)
(d) The GPUs must communicate/synchronise, e.g. share results/updated weights (gradients) after each step, and this communication overhead grows as more GPUs are added(1)
(d) Some parts of the work cannot be split / must be done in sequence (e.g. loading data, combining results), so they take the same time however many GPUs are used (accept Amdahl’s law; accept idle GPUs waiting for others/load imbalance)(1)
Answer: (a) 27 h (b) 10 (c) 62.5% (d) Communication/synchronisation overheads between GPUs and sequential parts of the work.
Examiner tip: Efficiency must be a percentage; giving 0.625 without converting loses the mark unless 62.5% is also seen.

10Stretch6 marks

A university research group needs GPUs to train machine-learning models for the next three years. It expects to use about 300 GPU-hours per month. It has two options:

Option A: buy its own GPU server for £36 000, with running costs (electricity, cooling, maintenance) of £200 per month.
Option B: rent equivalent GPU instances from a cloud provider at £4.50 per GPU-hour.

Evaluate the two options for the research group. [6]

Cost of A over 3 years: 36 000 + 36 × 200 = £43 200(1)
Cost of B over 3 years: 300 × 4.50 × 36 = £48 600 (£1350 per month), so at 300 hours/month A is about £5400 cheaper(1)
Break-even is about 267 GPU-hours per month ((1000 + 200) ÷ 4.50); if use falls below this, B is cheaper(1)
A: large upfront capital cost; group is responsible for maintenance, cooling and repairs; hardware may become outdated within three years; capacity is fixed, so large jobs queue(1)
A: data stays on site (privacy/security of sensitive research data); no need to upload large data sets; no dependence on an internet connection or provider(1)
B: no upfront cost, pay only for what is used; elastic/scalable, so many GPUs (or TPUs/newer hardware) can be rented for short bursts or deadlines; provider handles maintenance and upgrades(1)
B: costs can rise if usage grows or prices change; data must be transferred to a third party (privacy, bandwidth, transfer time); vendor lock-in(1)
Justified conclusion, e.g. A if usage is steady at or above ~267 hours/month or data is sensitive; B if usage is uncertain/bursty or the latest hardware is needed(1)

Level	Marks	What it looks like
Level 3	5–6	Balanced evaluation of both options covering cost (with correct use of the figures) and several non-cost factors such as scalability, maintenance, obsolescence and data security; reaches a clear conclusion justified by the evidence.
Level 2	3–4	Discusses advantages and disadvantages of both options with some development; may use the figures with minor errors or partially; conclusion is missing or only partly justified.
Level 1	1–2	Identifies some relevant points about one or both options with little explanation; any calculation is incorrect or unused; no conclusion.
Level 0	0	Nothing worthy of credit.

Answer: Over 3 years A costs £43 200 and B £48 600, so A is cheaper at 300 h/month (break-even ≈ 267 h/month), but B offers no upfront cost, scalability and maintained, up-to-date hardware; a justified conclusion depends on how steady usage is and how sensitive the data is.
Examiner tip: Top marks need a conclusion that is justified using the figures; listing advantages of each option without weighing them caps the answer in the middle band.

Stuck on this topic?

Book a free 30-minute session on WhatsApp and we'll find exactly which step is costing you the marks. Or build a timed paper on it.

Free 30-minute session