This book covers fourteen parts and seventy-two chapters, which is more than any single term can hold. This appendix carves five teachable tracks out of that material, each aimed at a different audience and calendar. Every track lists the chapters it uses, so you can assemble reading and lab schedules directly from the table of contents. The first two tracks are given as full week-by-week tables that map each session to book parts and chapters; the remaining three are given as compact chapter-to-week maps that you can expand the same way. A standard week is one topic lecture plus one graded lab, and every track ends with a project rather than a final exam, because a sensor system is only understood once it has been built and evaluated on data it has not seen.
Track 1: Undergraduate Sensory AI (14 weeks)
A first course for third-year or fourth-year students who have taken linear algebra, probability, and one machine-learning course. It walks the classical-to-neural arc for one-dimensional sensor streams and stops before foundation models. The project is a wearable activity recognizer evaluated leave-one-subject-out.
| Wk | Topic | Book | Lab |
|---|---|---|---|
| 1 | What sensory AI is; the sense-model-act loop | Ch 1 | Read a live IMU stream, plot and annotate it |
| 2 | Sensor physics and measurement models | Ch 2 | Fit a calibration curve to a noisy sensor |
| 3 | Signals, sampling, aliasing, synchronization | Ch 3 | Alias a signal on purpose, then fix the rate |
| 4 | Probability, estimation, uncertainty | Ch 4 | Maximum-likelihood fit with error bars |
| 5 | Data engineering and leakage-safe splits | Ch 5 | Build a subject-grouped train/test split |
| 6 | Filtering and denoising | Ch 6 | Low-pass and median filters on real motion data |
| 7 | Spectral and time-frequency analysis | Ch 7 | Spectrogram of gait; read cadence off it |
| 8 | Feature engineering and dimensionality reduction | Ch 8 | Handcrafted features plus PCA baseline |
| 9 | Bayesian filtering and the Kalman family | Ch 9 | 1D Kalman filter for a drifting sensor |
| 10 | Neural representations for sensor streams | Ch 13 | 1D CNN on windowed accelerometer data |
| 11 | Recurrent and temporal convolutional models | Ch 14 | Compare an LSTM and a TCN on the same task |
| 12 | Inertial sensors; human activity recognition | Ch 23, 26 | Full HAR pipeline on UCI-HAR |
| 13 | Uncertainty, calibration, conformal prediction | Ch 18 | Calibrate the HAR model; add conformal sets |
| 14 | Evaluation protocols; project presentations | Ch 65 | Leave-one-subject-out report and defense |
The leakage discipline of week 5 is the spine of the course: every later lab is graded partly on whether the split keeps each subject on one side of the boundary. Formally the model is scored on \(\mathbb{E}_{s \notin \mathcal{S}_{\text{train}}}[\ell]\), the expected loss over subjects never seen in training, which is the only number that predicts field behavior.
Track 2: Graduate Cyber-Physical and Foundation-Model Sensing (14 weeks)
A research-oriented graduate seminar assuming a solid deep-learning background. It pairs modern sequence architectures with foundation models and closes the loop on trust, cyber-physical security, and fleet operations. Each week combines a lecture, one paper the class critiques, and a lab; the project is an original benchmark or method on a leakage-safe cyber-physical dataset.
| Wk | Topic | Book | Lab |
|---|---|---|---|
| 1 | Leakage-safe datasets and evaluation, revisited hard | Ch 5, 65 | Break a published split; measure the inflation |
| 2 | Factor graphs and smoothing for state estimation | Ch 11 | Batch smoothing on a trajectory with a solver |
| 3 | Transformers for sensor data | Ch 15 | Patch-and-attend model on multivariate series |
| 4 | State-space models: S4, Mamba, linear-time sequences | Ch 16 | Reproduce a long-context SSM baseline |
| 5 | Self-supervised and contrastive sensor learning | Ch 17 | Pretrain with masking; probe the frozen encoder |
| 6 | Time-series foundation models | Ch 19 | Zero-shot forecast; guard against contamination |
| 7 | Sensor and wearable foundation models | Ch 20 | Fine-tune a sensor FM on a small labeled set |
| 8 | Multimodal sensor-language models and reasoning | Ch 21 | Caption and query a sensor window with an SLM |
| 9 | LLM agents for sensing and operations | Ch 22 | Agent that triages alarms from a telemetry feed |
| 10 | Deep multimodal fusion; missing-modality robustness | Ch 50 | Drop a modality at test time; measure the fall |
| 11 | Distribution shift, OOD, test-time adaptation | Ch 66 | Adapt a model to an unseen operating regime |
| 12 | Cyber-physical anomaly detection and ICS security | Ch 38 | Event-level anomaly scoring on SWaT or WADI |
| 13 | Robustness, sensor spoofing, functional safety | Ch 68 | Craft a spoof; test detector and safety envelope |
| 14 | MLOps for sensor fleets; project defense | Ch 69 | Ship the project with drift monitoring |
Week 12 is deliberate about metrics: on ICS datasets attacks appear only in the test window, so point-adjusted accuracy can turn one lucky detection into a near-perfect score. The seminar requires range-based scoring, and asks students to report both the naive and the honest number so the gap is visible.
Track 3: IoT and Edge AI Professional Course (7 weeks)
A short, hands-on course for working engineers who need to deploy models on constrained hardware. It skips most classical theory and drives toward a microcontroller demo. Assumes Python fluency and basic ML.
| Wk | Topic | Book |
|---|---|---|
| 1 | Sensory AI overview; sensor data engineering | Ch 1, 5 |
| 2 | Neural representations for sensor streams | Ch 13, 14 |
| 3 | Edge AI fundamentals and model optimization | Ch 59 |
| 4 | Streaming inference and online learning | Ch 60 |
| 5 | TinyML and microcontroller sensor AI | Ch 61 |
| 6 | On-device continual learning; batteryless sensing | Ch 62, 63 |
| 7 | Federated and privacy-preserving sensor AI; capstone | Ch 64 |
The capstone is a quantized keyword or gesture spotter running on a microcontroller under a fixed memory and latency budget, reported as an accuracy-versus-\(\text{flash}\times\text{latency}\) frontier rather than a single accuracy number.
Track 4: Robotics Sensor Fusion and Embodied AI Seminar (14 weeks)
A graduate seminar for robotics students, pairing geometric perception with learned world models and embodied policies. It reads roughly one part per fortnight and builds toward a fusion-and-navigation project on a real or simulated platform.
| Wk | Topic | Book |
|---|---|---|
| 1 | Orientation estimation and dead reckoning | Ch 24 |
| 2 | Localization, GNSS, RF positioning | Ch 25 |
| 3 | Depth and 3D sensing fundamentals | Ch 40 |
| 4 | Point clouds and LiDAR AI | Ch 42 |
| 5 | BEV perception and 3D occupancy | Ch 43 |
| 6 | Foundations of sensor fusion | Ch 48 |
| 7 | Probabilistic and Bayesian fusion | Ch 49 |
| 8 | Neural fields and Gaussian splatting for sensing | Ch 51 |
| 9 | SLAM and spatial AI | Ch 52 |
| 10 | World models and predictive sensing | Ch 53 |
| 11 | Tactile sensing and electronic skin | Ch 56 |
| 12 | Proprioception, exteroception, robot perception | Ch 57 |
| 13 | Vision-language-action and embodied foundation models | Ch 58 |
| 14 | Integration and project defense | Ch 71, 72 |
Track 5: Health and Wearable AI Short Course (7 weeks)
A short course for biomedical or data-science students focused on biosignals and the regulatory reality that surrounds them. Subject-independent evaluation is non-negotiable throughout, and the project is a wearable cardiovascular estimator with a validation and privacy write-up.
| Wk | Topic | Book |
|---|---|---|
| 1 | Biosignal foundations; signals and sampling | Ch 3, 28 |
| 2 | ECG and cardiac AI | Ch 29 |
| 3 | PPG and wearable cardiovascular sensing | Ch 30 |
| 4 | EMG and neuromuscular AI | Ch 32 |
| 5 | Continuous and contactless health monitoring | Ch 33 |
| 6 | Uncertainty, calibration, conformal prediction | Ch 18 |
| 7 | Clinical validation, regulation, biometric privacy; capstone | Ch 34 |
Across every track, three rules travel with the course and should appear on the first slide of each. Split on the group, never on the window, so the reported number measures generalization and not memory. Report uncertainty alongside every point estimate, because a sensor decision without a confidence is not actionable. And evaluate on the shift you actually care about, whether that shift is a new subject, a new machine, a new room, or a new day. A student who leaves with those three habits has learned the part of sensory AI that outlasts any particular architecture.