class 09 · unit I · challenges, opportunities, recap
What is still hard, and where it is going
Eight classes made edge AI sound possible. Today we are honest about what makes it hard, look at where the field is heading, and pull all of Unit I into one picture.
Plan: 20 min challenges, 12 min opportunities, 13 min recap, 15 min quiz.
last class's exit ticket
A 300 Hz fault, sampled at 400 Hz.
Sample faster than 600 Hz to see it honestly. At 400 Hz it masquerades as 100 Hz. A design decision in Class 8 becomes a safety problem in the field. That is today's theme.
Use it as the bridge: wrong sampling is a challenge nobody notices until something breaks.
Every challenge falls into one of four families.
The "C" tags point back to the class where students first met each problem. Ask: which family worries you most for your Class 3 product?
resources
On a battery, every design choice is a lifetime.
Same battery, three designs. Illustrative orders of magnitude, not measurements; the log scale hides how big the gaps are.
The cheapest milliwatt is the one you never spend: sleep most of the time, wake on something small (Class 5's cascade), and send results rarely (Class 2's B).
Ask what a doorbell on batteries would have to give up. Then: why do many real video doorbells need wiring? (Exactly this chart.)
data
Most edge AI projects fail on data, not on models.
scarce
no dataset exists
nobody has published "vibration of this pump, failing". You collect it yourself.
unlabelled
labels cost people
hours of audio, and someone must mark every "open"
skewed
lab ≠ field
quiet room in training, traffic noise at the door
drifting
the world moves
worn motors, new accents, a new paint colour on the wall
Unit III's feasibility checks start here: do we have, or can we get, the data? If not, stop before building anything.
Ask each team: for your Class 3 product, where would the training data come from? Most will realise they would have to record it themselves.
engineering
A bad update can break ten thousand devices at once.
Each square is a device in the field. A new model is ready. Do you send it to all of them?
Staged rollout: send it to 10% first. One of them misbehaves (red). Stop, fix, try again: 90% of the fleet never noticed.
Also hard: every chip family needs its own build (C4), devices may be offline for weeks, and a failed update must never leave a device unable to boot. Plan the update path on day one (C5).
Ask what a failed update looks like for a doorbell versus an insulin pump. The second is why medical devices have such strict update rules.
trust
The device is in someone else's hands.
Physical access
security
Anyone can buy one, open it and read its flash. Your model, and any secrets in the firmware, must be protected as if the attacker holds the board. Because they can.
Who is it for?
fairness
A wake word trained mostly on one accent works worse for everyone else. On the edge, nobody sees the failures: they happen silently, in someone's home.
Who answers?
responsibility
The fall detector missed a fall. The lock opened for a stranger. Who is responsible: the model, the maker, the data?
This is the "Responsible AI in the Edge AI workflow" topic in Unit III, and "Evaluation and Responsible AI" in Unit V.
Keep it as questions, not lectures. One good discussion prompt: should a doorbell's person detector be allowed to recognise faces? Who decides?
Where the field is going.
Specialised chips
NPUs in phones today, in microcontrollers next: more inferences per milliwatt
→ Unit II
Better compression
quantisation, pruning, distillation keep shrinking what a model needs
→ Unit IV
Learning on the device
adapting to one user's voice or one machine's normal, without the cloud
→ Unit III
Federated learning
many devices improve a shared model; raw data stays home
→ next slide
Small language models
compact language models already run on phones; very small ones on boards with extra RAM
→ beyond the syllabus
India-shaped problems
uneven connectivity makes R the strongest letter: farms, clinics, rail, forests
→ your projects
Each card points to where it comes back in this course. The last card is deliberate: the strongest BLERP letter in India is often R, and that is a design opportunity, not just a constraint.
Federated learning: the data stays home, the learning travels.
Used for next-word prediction on phone keyboards. Ask which BLERP letter it serves (P, and B). The open problems: devices that are offline, slow or malicious.
Unit I in one picture.
Before each tag appears, ask the room where that class's idea lands on the doorbell. Every class from 1 to 8 should find a place; Class 9 is everything that can go wrong with it.
Unit I in nine sentences.
class
if you remember one thing
class 1
Edge AI runs inference on or near the device that collects the data.
Every product is a sensor, a decision, and a reason to decide locally.
class 4
Five parts, three memories, seven software layers; the model is one of seven.
class 5
Sense, prepare, infer, decide, act; five patterns for splitting the work.
class 6
Data is a stream; every stage of the pipeline is a memory budget.
class 7
A model is weights plus multiply-adds; weights fit flash, activations must fit SRAM.
class 8
DSP turns 16,000 numbers into 1,960 better ones; sample above twice the highest frequency.
class 9
Resources, data, engineering, trust: the four families of what is still hard.
Ask students to say each sentence before it appears. This is the revision sheet for the CIE: tell them to photograph it.
CIE quiz · vote with hands, then click
Eight questions, one per idea from the unit. For a written CIE, use the same questions on paper; the deck version is for a class vote. The score counts first clicks only.
what comes next
Unit I treated the board as a box. Unit II opens it.
sensors, signals and sources of data
acoustic and vibrationvisual and scenemotion and positionforce and tactileoptical, electromagnetic, radiationenvironmental, biological, chemical
processors for edge AI
microcontrollers and DSPssystem-on-chipdeep learning acceleratorsFPGAs and ASICsedge serversmulti-device architectures
Every one of these was a single line in Class 4. Each gets real time now.
The chips are the prescribed Unit II topics, in the syllabus's own wording.
exit ticket · end of Unit I
You are shipping the doorbell next month. Which challenge from today worries you most, and what is one thing you would do to reduce it?
End of Unit I. Nine classes, one doorbell, and a vocabulary for everything that follows.
Good answers name a family (resources, data, engineering, trust) and a concrete action: a staged rollout, recording training audio at real doors, a battery budget, encrypting the firmware.