FTM 26WT: Examiner Notes
Confidential. Not published. This file sits in a gitignored notes/ folder and is excluded from the Quarto render list, so it never reaches docs/ or the public site.
Oral examination, 20.01.2027 from 14:00. Eight candidates, 8 min presentation plus 8 min Q&A. Grading 30 % presentation / 70 % Q&A.
1. The theory being tested
Emergence (from the Introduction unit)
Working definition used in class: emerging technologies are already under development or in early adoption and have the potential to significantly alter the status quo.
Rotolo, Hicks & Martin (2015) attributes, which students must be able to apply:
| # | Attribute | What to probe |
|---|---|---|
| 1 | Radical novelty | Is the novelty in the artefact, or only in the application? |
| 2 | Relatively fast growth | Growth in what: publications, patents, deployments, revenue? |
| 3 | Coherence | Is there a recognisable community and a converging approach? |
| 4 | Prominent impact | Impact on whom, and is it demonstrated or projected? |
| 5 | Uncertainty and ambiguity | What could make this fail, and who disagrees about its meaning? |
Attributes 2 and 5 separate strong candidates from weak ones. Weak candidates treat all five as a checklist to be ticked; strong candidates argue where their technology fails an attribute.
The four hypotheses
| # | Hypothesis | Unit |
|---|---|---|
| H1 | Emerging information technologies enable multimodal and immersive systems | Multimodality and Immersion |
| H2 | Emerging information technologies enable intelligent and affective systems | Intelligence and Affection |
| H3 | Emerging information technologies enable strongly interconnected and distributed systems | Interconnectedness and Distributivity |
| H4 | Emerging information technologies enable mobile and ubiquitous systems | Mobility and Ubiquity |
Assignment and mapping
| ID | Technology | Primary | Cross-examine on |
|---|---|---|---|
| 410335 | Light-Field Displays | H1 | H4, H3 |
| 298791 | Mid-Air Haptics | H1 | H4, H2 |
| 373533 | Neuromorphic Computing | H2 | H4, H1 |
| 341501 | Federated Learning | H2 | H3 (very strong) |
| 360948 | Quantum Key Distribution | H3 | H4; H2 as a deliberate misfit |
| 344669 | DePIN | H3 | H4; H2 as a deliberate misfit |
| 410629 | E-Textiles | H4 | H2 (strong), H1 |
| 369420 | Batteryless IoT Devices | H4 | H3, H2 |
Students were not told their mapping and were instructed that all four hypotheses are examinable. Do not reveal the mapping during the examination.
2. Generic questions (usable for any candidate)
Keep these in reserve for a candidate who stalls, or to fill time.
- Which of Rotolo’s five attributes does your technology satisfy least well, and why? Good answer: names one, argues it, does not retreat into marketing claims.
- Who would have to change their behaviour for this technology to have prominent impact? Good answer: identifies a concrete actor and an adoption barrier, not “society”.
- Your technology sits inside one hypothesis comfortably. Name the hypothesis it fits worst, and say what would have to be true for it to fit.
- Distinguish the technology from the system built on it. Which of the two does the hypothesis actually describe? This one catches the most common confusion across all eight topics.
- If I gave this to you five years from now, which attribute would have changed most?
Timing
Eight minutes allows roughly four questions with follow-ups. Suggested shape: one emergence question, one primary-hypothesis question, one cross-hypothesis question, one application or ethics question. Do not spend more than two minutes on any single question.
Red flags across all candidates
- Reciting the five attributes without applying them.
- Claiming all four hypotheses apply equally. They never do.
- Describing the vendor roadmap instead of the technology.
- Confusing a demonstration with deployment when arguing “fast growth”.
3. Per technology
410335 — Light-Field Displays (H1)
Characterization. A light-field display reproduces radiance as a function of both position and direction, so different images reach different eye positions. This yields motion parallax and, in principle, correct accommodation cues, which conventional stereoscopic 3D cannot deliver. Implementations include integral imaging with microlens arrays, stacked and compressively driven LCD layers, and multi-projector arrays. The governing constraint is the spatio-angular resolution tradeoff: angular resolution is bought with spatial resolution and with very large data rates. Maturity is niche commercial (Looking Glass, automotive HUDs, medical visualisation) rather than consumer scale.
Q1 (emergence). Light-field capture and display were described by Lippmann in 1908 and revived by Levoy and Hanrahan in 1996. On what grounds can you still call this an emerging technology? Solution notes: The strong answer separates the optical principle (old) from the enabling substrate (new): panel density, compute for real-time light-field synthesis, and content pipelines. Radical novelty should then be located in the system, not the concept. A candidate who simply asserts novelty has failed the question. Accept a well-argued answer that concedes radical novelty is weak here and shifts the claim to attribute 2 or 4.
Q2 (H1). The hypothesis says emerging technologies enable multimodal and immersive systems. Light-field displays address only vision. Justify the immersion claim. Solution notes: Expect the accommodation-vergence conflict: stereoscopic displays force the eye to focus at the screen while converging at the virtual depth, causing fatigue and breaking presence. Light-field displays can resolve this. Strong answers connect to presence and to the glasses-free property, which removes the encumbrance that a headset imposes. Multimodality is genuinely weak here, and saying so is correct.
Q3 (cross, H4). Place this on Milgram and Kishino’s virtuality continuum, then argue whether it pushes towards ubiquity (H4) or away from it. Solution notes: Light-field displays are environment-fixed, so they sit nearer the real end and support augmented reality without worn hardware. The H4 tension is productive: removing the headset makes the technology more ubiquitous in Weiser’s sense (the interface disappears), but the display itself is bulky, power-hungry and fixed, which works against mobility. A good candidate holds both halves.
Q4 (application and ethics). Who benefits first, and what is the equity implication? Solution notes: Medical imaging, design review, air-traffic and automotive HUDs, where depth judgement is safety-relevant, precede entertainment. Ethics: cost concentrates access; shared public displays raise consent questions because imagery can be made visible to one viewer position and not another, which enables targeted display without the bystander knowing.
298791 — Mid-Air Haptics (H1)
Characterization. Tactile sensation delivered without contact, produced by a phased array of ultrasonic transducers that focuses acoustic radiation pressure onto the skin. The focal point is modulated, by amplitude modulation or spatiotemporal modulation, so that mechanoreceptors can detect it. Forces are small, on the order of millinewtons, so the technology conveys shape, texture and confirmation cues rather than resistance. Commercially live in niches: automotive infotainment, digital signage, and sterile medical interaction (Ultraleap is the dominant vendor).
Q1 (emergence). Apply attribute 3, coherence. Is there a coherent research community here, and what follows if coherence rests on a single vendor? Solution notes: Coherence is high technically, since nearly all work uses focused ultrasound phased arrays, but it is fragile: one dominant vendor and a small academic community mean the trajectory is tied to one firm’s survival. Strong answers notice that high coherence can be a sign of a narrow field rather than a mature one, and link this to attribute 5.
Q2 (H1). This is the clearest multimodality case in the cohort. Explain why adding touch changes the system qualitatively, not just quantitatively. Solution notes: Expect the argument that visual and auditory channels are already saturated, and that haptic confirmation closes the perception-action loop, allowing eyes-free interaction. In cars this is the actual driver of adoption: a touchscreen requires visual attention, a mid-air gesture with haptic confirmation does not. Reference to Suh and Prophet (2018) on immersive technology and sensory channels is a plus.
Q3 (cross, H4). Argue mid-air haptics as a ubiquity technology. Solution notes: Touchless public interfaces, hygiene in shared and clinical settings, and interaction without a worn or held device. Weiser’s criterion is that the technology recedes: here the interface has no physical presence at all, which is a strong fit. Counterargument worth drawing out: the transducer array is a conspicuous, mains-powered, directional device with a small interaction volume, so it is ubiquitous only within a fixed envelope.
Q4 (application and ethics). What are the limits, and is there a safety question? Solution notes: Force output limits it to confirmation and light texture, not force feedback; the array requires line of sight to the hand; the working volume is small. Safety: continuous ultrasound exposure, audible subharmonics, and hearing concerns for people and animals. A candidate who claims it can simulate solid objects has not understood the force budget.
373533 — Neuromorphic Computing (H2)
Characterization. Hardware architectures that imitate the structure and dynamics of neural systems: spiking neurons, event-driven asynchronous operation, and memory co-located with compute, which sidesteps the von Neumann bottleneck. Representative silicon includes Intel Loihi and Loihi 2, IBM TrueNorth and NorthPole, SpiNNaker and BrainScaleS, often paired with event cameras. The appeal is energy per inference on sparse, temporally structured workloads. The barrier is software: no mature programming model, no settled benchmarks, and limited tooling compared with conventional deep learning stacks.
Q1 (emergence). Carver Mead coined “neuromorphic” in 1990. Attribute 2 is relatively fast growth. Does this technology satisfy it? Solution notes: This is the sharpest emergence question in the cohort and should be pressed. An honest answer concedes that by publication and deployment measures growth has been slow and episodic for three decades, and then argues either that the relevant growth is recent and hardware-driven, or that the technology fails attribute 2 and is better described as persistent rather than emerging. Both are creditable. Uncritical assertion of fast growth is not.
Q2 (H2). The hypothesis is that emerging technologies enable intelligent and affective systems. Neuromorphic chips run the same classification tasks as a GPU. Where is the enablement? Solution notes: The key distinction is substrate versus capability. Neuromorphic computing does not make systems more intelligent in the functional sense; it changes the power, latency and form-factor envelope in which intelligence can be deployed, which enables always-on local intelligence that would otherwise be impossible. Candidates who claim neuromorphic chips are inherently “more intelligent” or “closer to the brain” in a cognitive sense should be pushed to defend it and will usually not be able to.
Q3 (cross, H4). Argue the ubiquity case. Solution notes: Microwatt-scale always-on sensing permits intelligence in devices that cannot host a GPU or a battery of any size, which links directly to batteryless and wearable contexts. This is the strongest practical case for the technology and connects to candidate 369420. A good answer also notes the affective link: continuous low-power physiological sensing is what makes affect recognition feasible outside the lab.
Q4 (application and ethics). What follows ethically from moving inference onto the device? Solution notes: Local processing is privacy-favourable because raw data need not leave the device, but it also makes continuous sensing economically viable, which expands surveillance capacity. The candidate should hold the tension rather than resolve it in one direction.
341501 — Federated Learning (H2)
Characterization. A distributed training paradigm in which many clients train on local data and exchange only model updates, which a coordinator aggregates, so raw data never leaves the holder. The canonical algorithm is FedAvg (McMahan et al., 2017). Two regimes differ sharply: cross-device, with millions of unreliable clients and non-IID data, and cross-silo, with a few institutional participants under regulatory constraint. Live at scale in mobile keyboards and on-device personalisation; cross-silo use in healthcare and finance is the emerging frontier. Privacy is not automatic: gradient-inversion attacks can reconstruct training data, so differential privacy and secure aggregation are required additions.
Q1 (emergence). Distinguish radical novelty from recombination here. Solution notes: Distributed optimisation and parameter servers predate it; the novelty is the inversion of the data-gravity assumption, moving computation to data for institutional rather than performance reasons. Strong answers argue that the novelty is architectural and regulatory rather than algorithmic. Conceding limited radical novelty while defending attributes 2 and 4 is a good answer.
Q2 (H2 and H3 together). You were assigned this under intelligence, but the mechanism is distribution. Which hypothesis does federated learning actually evidence? Solution notes: This is the best cross-hypothesis case in the cohort and both answers are defensible if argued. H2: it enables intelligent systems in domains where data cannot be centralised, so capability is the outcome. H3: the distinctive content is the distributed architecture and the coordination protocol, so H3 describes the mechanism while H2 describes the effect. The ideal answer separates mechanism from effect explicitly. Press anyone who picks one without acknowledging the other.
Q3 (critical). Does federated learning make a system private? Solution notes: No, and the candidate must say so. Model updates leak information; gradient inversion can recover images and text. Privacy requires differential privacy, secure aggregation, or both, each with a utility cost. A candidate who equates “data stays local” with “privacy preserved” has the central misconception of this topic and should be marked down.
Q4 (application and ethics). Who is disadvantaged by non-IID data? Solution notes: Clients whose distribution differs from the majority are underserved by the global model, so federation can entrench majority bias while appearing egalitarian because everyone contributes. Good answers mention personalisation layers or client clustering as mitigations, and note that fairness and privacy budgets compete.
360948 — Quantum Key Distribution (H3)
Characterization. A method of establishing a shared secret key between two parties whose security rests on quantum mechanics rather than computational hardness: measurement disturbs the state and unknown states cannot be cloned, so eavesdropping is detectable. BB84 (Bennett & Brassard, 1984) is the canonical prepare-and-measure protocol; E91 is entanglement-based. Practical limits dominate: fibre loss caps direct links at a few hundred kilometres, trusted relay nodes reintroduce a classical trust assumption, satellite links (Micius) extend reach, and quantum repeaters remain immature. Commercial equipment exists and metropolitan networks are deployed, most extensively in China.
Q1 (emergence, attribute 4). Post-quantum cryptography is software-only, standardised, and deployable on existing hardware. Given that, can QKD claim prominent impact? Solution notes: This is the decisive question for this candidate. A strong answer concedes that PQC is the likely default for most of the internet and confines QKD’s prominent impact to settings with long-term secrecy requirements and physical control of the link, such as state, defence and inter-datacentre backbones. Candidates should know that several national security agencies actively recommend PQC over QKD. Asserting that QKD will replace classical key exchange is wrong and should be challenged.
Q2 (H3). The hypothesis concerns interconnected and distributed systems. Trusted-node QKD requires every relay to be trusted. Is that distributed or centralised? Solution notes: It is a centralising architecture masquerading as a distributed one: security degrades to the weakest trusted node, so the trust model is hub-like even though the topology is networked. Strong answers distinguish topological distribution from trust distribution, and note that genuine end-to-end distribution awaits quantum repeaters.
Q3 (cross, H2 as misfit). Make the case that QKD relates to H2. Then tell me why it fails. Solution notes: The point of the question is to reward an honest negative. There is no meaningful intelligence or affect dimension; the only tenuous link is machine-learning-based side-channel detection on QKD hardware, which is incidental. A candidate who manufactures a strong H2 connection is overreaching and should be marked down; one who argues clearly that the hypothesis does not apply has understood that the hypotheses are claims, not categories.
Q4 (application and ethics). What is the security risk introduced by the implementation rather than the theory? Solution notes: Information-theoretic security is a property of the protocol, not the device. Detector blinding, photon-number-splitting and other side channels attack the hardware, which is why decoy-state and measurement-device-independent variants exist. The ethical and policy angle is that “unbreakable” marketing claims mislead procurement.
344669 — Decentralized Physical Infrastructure Networks (H3)
Characterization. Networks in which individuals deploy and operate physical infrastructure and are compensated in tokens, with on-chain verification that the work was really performed. Examples span wireless coverage (Helium), storage (Filecoin, Arweave), GPU compute (Render), mapping (Hivemapper) and environmental sensing (WeatherXM). The mechanism of interest is economic: token incentives solve the cold-start problem that blocks infrastructure build-out, by paying the supply side before demand exists. The recurring weakness is precisely there: supply frequently outruns real demand, and participation tracks token price more closely than service revenue.
Q1 (emergence). Where is the radical novelty: in the hardware or in the coordination mechanism? Solution notes: The hardware is commodity. Novelty is organisational and economic, namely token-incentivised coordination with proof of physical work. Strong answers argue that Rotolo’s framework tolerates this, because a technology can be a sociotechnical arrangement, and then ask whether an economic mechanism should count as a technology at all. That reflection is a distinction-level answer.
Q2 (attribute 5). DePIN is the strongest case in this cohort for uncertainty and ambiguity. Develop that. Solution notes: Expect: regulatory ambiguity over whether tokens are securities; definitional ambiguity about what counts as DePIN; and genuine uncertainty about whether the model survives a token bear market. Helium is the instructive case, with extensive coverage built and persistently weak paid demand. A candidate who treats growth in deployed nodes as evidence of impact should be asked who is paying for the service.
Q3 (H3). Which kind of distribution does this evidence, technical or organisational? Solution notes: Primarily organisational, that is ownership and operation are distributed, while the technical stack often retains centralised components such as the chain, the oracle, or a single vendor’s gateway. The best answers notice that DePIN distributes capital expenditure and risk rather than computation, which is a different claim from the one H3 usually carries.
Q4 (application and ethics). What happens to participants when the token price falls? Solution notes: Individuals bear the capital cost of hardware bought on the expectation of token income, so downside risk is pushed onto non-professional participants who may not understand it. Service reliability also degrades when operators leave, which matters if anyone depends on the coverage. Expect some discussion of whether this is infrastructure or speculation.
410629 — E-Textiles (H4)
Characterization. Textiles with electronic function integrated at the fibre or fabric level: conductive yarns for interconnect, knitted or woven capacitive, resistive and piezoresistive sensors, printed electronics on fabric, and occasionally on-garment energy harvesting. The hard problems are not electrical but textile: washability, mechanical durability over thousands of flex cycles, reliable rigid-to-flexible connectors, power, and manufacture at textile cost and scale. The viable commercial niches are medical and sports monitoring; consumer fashion applications have largely failed, Google and Levi’s Project Jacquard being the cautionary case.
Q1 (emergence). Conductive textiles are centuries old and the modern field dates to the late 1990s. Which attributes does this technology fail? Solution notes: It fails attribute 2 fairly clearly, since growth has been slow and punctuated by abandoned consumer products, and it is weak on attribute 3, because the field is fragmented across textiles, materials science and HCI with no converging approach. A candidate who can name two failing attributes and still argue the technology is worth attention has given a strong answer. This topic rewards honesty more than enthusiasm.
Q2 (H4). Connect this to Weiser’s account of ubiquitous computing. Solution notes: Weiser’s criterion is that the most profound technologies disappear and weave themselves into everyday life. Clothing is the limiting case of a disappearing interface: already worn, already everywhere, requiring no new object. The candidate should then confront the gap between that vision and the record of commercial failure, and identify why: cost, laundering, and the absence of an application that justifies replacing a garment.
Q3 (cross, H2). Argue the affective computing case. Solution notes: This is the strongest cross-link in the cohort and should be pursued. Garments sit against the skin continuously, which gives access to heart rate, heart-rate variability, respiration, electrodermal activity and posture, that is exactly the physiological substrate for affect recognition. Dellermann et al. (2019) on hybrid intelligence is a reasonable anchor if the candidate wants to argue the human-machine complementarity angle. Expect acknowledgement that inference from physiology to emotion is contested.
Q4 (application and ethics). What is specific about the privacy question for clothing? Solution notes: Continuous, involuntary and intimate sensing, with no natural off switch and unclear consent when a garment is shared, gifted, resold or required by an employer. Workplace mandates are the sharpest case. Good answers raise the inference problem: innocuous signals combine to reveal health conditions, pregnancy, or stress, which the wearer never disclosed.
369420 — Batteryless IoT Devices (H4)
Characterization. Devices that sense, compute and communicate without a battery, drawing ambient energy from radio frequency, light, thermal gradients or vibration into a small capacitor. Two consequences define the field. First, execution is intermittent, since power fails constantly, which requires checkpointing and task-based programming models so computation can survive repeated power loss. Second, communication commonly uses backscatter, reflecting an ambient or dedicated carrier rather than generating one, which cuts transmit energy by orders of magnitude. RFID and NFC-powered sensing are the mature end; general-purpose batteryless compute remains a research field with early commercial entrants.
Q1 (emergence). Locate the radical novelty precisely. Solution notes: Energy harvesting is old and RFID is mature, so the novelty is the programming model. Intermittent computing breaks the assumption of forward progress that essentially all software assumes, which is a genuine change in abstraction rather than an efficiency gain. A candidate who identifies intermittency as the novel element has understood the topic; one who talks only about harvesting efficiency has not.
Q2 (H4). Why are batteries the binding constraint on Weiser’s vision? Solution notes: Expect an argument from arithmetic: ubiquity implies very large numbers of devices, and any per-device maintenance event, multiplied by that number, is economically and logistically impossible. Batteries also dictate size, cost and disposal. Removing them is therefore not an optimisation but a precondition for deployment densities that H4 presumes. This is the cleanest hypothesis argument available to any candidate in the cohort.
Q3 (cross, H3). How does intermittency change what the network must tolerate? Solution notes: Nodes are unreachable most of the time and unsynchronised, so protocols must be asynchronous and delay-tolerant, and the network cannot assume a device will be awake to acknowledge. Backscatter also makes a node dependent on an external carrier, which means connectivity is a property of the environment rather than of the device. Strong answers note that this inverts the usual assumption of an autonomous networked node.
Q4 (application and ethics). Is this a sustainability win? Solution notes: Partly. It removes battery chemistry, replacement labour and the associated waste stream, which is real. Against that, it makes deployment so cheap that vastly more devices are deployed, each still containing semiconductors and rare materials, and typically with no end-of-life plan since nobody retrieves a sensor that needs no servicing. The strongest answers reach the rebound effect without prompting: efficiency enabling expansion that cancels the saving.
4. Marking guidance
Because the Q&A carries 70 %, the distinguishing behaviours are:
| Level | What it looks like in the Q&A |
|---|---|
| Very good | Applies the five attributes and names where the technology fails one; separates mechanism from effect; argues a hypothesis misfit honestly; uses lecture terminology precisely without prompting |
| Good | Correct terminology and a sound primary-hypothesis argument; needs a prompt to reach the critical or cross-hypothesis point |
| Satisfactory | Describes the technology accurately but treats the hypotheses as labels to attach rather than claims to test |
| Sufficient | Recalls the attributes and hypotheses on request, cannot apply them to the specific technology without substantial help |
| Not sufficient | Cannot connect the technology to the course theory, or defends claims the evidence contradicts after correction |
The single most informative probe across all eight candidates is the misfit question: asking a student to argue a hypothesis that does not apply and then explain why. It cannot be prepared by rote and it separates understanding from recall.