Dataset and labelling plan

1. Define three gestures plus an explicit unknown class
2. Collect from multiple users and orientations
3. Use equal capture duration per class
4. Keep test sessions separate from training captures
5. Document sample rate, window length and overlap

Training and validation pipeline

1. Inspect raw signals for clipping and missing samples
2. Extract spectral or statistical features consistently
3. Train a small classifier and review confusion matrix
4. Quantise to int8 and compare accuracy
5. Export a C++ library or .tflite model
6. Test unseen gestures and low-confidence rejection

Model deployment procedure

1. Freeze preprocessing and class order
2. Add model bytes and resolver to firmware
3. Allocate tensor arena from measured requirement plus margin
4. Match device sampling/window logic to training exactly
5. Run inference outside time-critical interrupt handlers
6. Log class, confidence, latency and rejection result

Memory and performance targets

Deployment checkpoints

// Pseudocode: identical preprocessing is mandatory
read_window(accel_buffer);
features = preprocess(accel_buffer);
if (invoke_model(features) != OK) enter_safe_fallback();
result = top_class_with_confidence();
if (result.confidence < 0.70) report_unknown();
else report_class(result);

Troubleshooting

Safety, privacy and model limits

• Do not collect identifiable motion data without consent.
• Remove names and unnecessary metadata from exported datasets.
• Gesture output must not directly control hazardous machinery.
• Model confidence is not calibrated safety probability.

Interview and viva questions

1. Why must one recording session not be split across train and test?
2. What changes during int8 quantisation?
3. Why is an unknown/reject policy necessary?
4. How do window length and overlap affect latency?
5. Which memory regions hold model, tensor arena and stack?

Lesson notes

Responsible TinyML workflow

Keep raw data, preprocessing, model version, compiler options and measured device results together. A desktop accuracy score does not prove embedded performance. Validate representative unseen samples on the actual target and define an explicit fallback for low confidence or out-of-distribution input.

Deployment evidence

Record the dataset split, confusion matrix, exported model hash, firmware build, board revision, peak memory, average and worst-case latency, power conditions and failures. Never treat an educational classifier as a safety-certified decision system.