Course overview

AOverview

This lecture is the course in miniature. It opens with the instructor's lab and six field problems his group has worked on, from a drone mapping a rail corridor to a gateway spotting intrusions, and it closes with five promises about what you will be able to do by the end of term. Every later lecture is one of those promises made concrete.

As a PhD student you need it for three reasons. The exams will ask you to expand SERV, place a technique in the five-step loop, and compute whether a model fits a device from its parameter count and bit width. The research project, whatever Blackboard finally calls it, will be a hardware Deployment under a budget of memory, compute, latency and energy, and this lecture is where you learn to see every problem as that budget. Beyond the course, any Edge AI system you design is judged by the same arithmetic: does the model fit, does it answer in time, and is it still right when the world is wet, noisy or hostile?

The path has five stops. You meet the lab and read its name, Secure, Efficient, Robust Vision, as a syllabus. You look at drones, farm robots, vehicles and IoT gateways and see the same problem each time: a model next to a sensor with no room to spare. You learn the question that unites them, how much model can this platform afford. You get the five takeaways and turn them into a working loop. And you plan week one, because a project that starts without a baseline never really starts.

Success looks like

• Expand SERV and attach one course theme and one lab project to each letter.
• Read any detection figure critically: separate parameters, FLOPs, latency and mAP, and treat a saliency map as attribution rather than proof.
• Explain with numbers why raw sensor data cannot be shipped to the cloud and why a moving vehicle sets the inference deadline.
• State the five takeaways in loop order and place any technique on the correct step.
• Compute weight storage as parameters times bits over eight and decide whether a model fits a named board.
• Distinguish on-device inference from on-device training by what each must hold in memory.
• Leave week one with a device tier, a candidate model and a baseline metrics log.

How the course is structured

Slide 11 gives the structure without a timetable. The five takeaways, measure efficiency, accelerate, analyse trade-offs, track the frontier, build and deploy, are the verbs of one loop, and the lectures posted so far follow it in order. The far end of the arc is a hedged reading: the takeaways echo MIT's 6.5940 almost word for word, and that course runs from metrics through pruning, quantization, architecture search, distillation and MCUNet to deployment and on-device training, so expect the same shape here, adjusted to what Blackboard announces.

How the course is graded

One calibration point is available. The MIT course these takeaways mirror grades entirely on five labs and a group project, with no tests or exams. That is MIT's scheme, not KFUPM's, and a graduate course here typically keeps exams. Read it as a signal about emphasis rather than a prediction: hands-on compression and deployment carry real weight, so the project deserves planning from the first week.

How to plan for the project

Takeaway 5 promises hands-on experience compressing and implementing models on resource-constrained devices. That is a hardware deployment with a report, and the students who struggle with it are rarely short of techniques. They are short of a baseline to compare against and a board that arrives on time.

1. Choose a device tier before a model. A 256 KB SRAM microcontroller, a Raspberry Pi and a Jetson module admit very different candidates, and the tier is an input to model selection, not an output of it.
2. Pick a candidate model that already fits or is one or two optimizations from fitting, and record its parameter count and float32 size on day one.
3. Start a metrics log before the first optimization: accuracy, size, peak memory, latency and energy, with one row per experiment and a note of what changed. This is the compare step made permanent.
4. Tie the project to a SERV theme. A model that is efficient and also robust to weather or shift, or lightweight and also secure against crafted inputs, is the kind of contribution the lab publishes.
5. Order the hardware in week one. Shipping time is the most common reason a deployment slips.

How to study this lecture

1. Read the full guide once for the big picture, then return to one part at a time when you want the numbers.
2. Answer every recall prompt in your head, or on paper, before you reveal it. Retrieval effort is what makes SERV, the loop and the size formula stick.
3. Take each quiz and read the explanation even when you are right.
4. Use the confidence-threshold and model-fit simulators until you can predict their output before you move the slider.
5. Mark a concept as understood only when you could explain it to a classmate without looking. Unmarked concepts show you where to return.
6. Come back after a few days and retry the prompts and quizzes cold. Spaced practice builds the long-term memory an exam needs.

BThe 2 parts

1. 01The instructor and edge AI in practiceWho teaches the course, and what his lab's projects show about running vision and security models on UAVs, farm robots, vehicles and IoT devices.5 conceptsSlides 1-830 min
   1. 1.1Know the instructor: SERV as the map of the course
   2. 1.2Multi-sensor payloads: what an inspection rig carries and why it cannot carry much
   3. 1.3Nano detectors on farms and in the rain: efficiency and robustness measured
   4. 1.4PNet-IDS: a small CNN that watches IoT traffic at line rate
   5. 1.5Prior works: safety-critical inference at the roadside and under the train
2. 02Course takeaways and getting startedWhat the course promises you will be able to do, the five pillars of efficient deep learning on constrained devices, and what to do in week one.3 conceptsSlides 9-1218 min
   1. 2.1Joining the cohort: introductions and the pre-course survey
   2. 2.2Five takeaways that form one optimization loop
   3. 2.3The syllabus lives on Blackboard: plan from week one

CGlossary and reference

• GlossaryEvery term from the lecture, defined once, with links back to the slides that use it.
• Reference sheetThe lecture on one printable page, for revision before a quiz or exam.