Edge AI
Edge AI is the practice of running artificial intelligence on the devices where data is created:microcontrollers, sensors, cameras, and small edge computers rather than sending raw data to distant servers. The device senses → interprets → decides → acts locally, in milliseconds, with tight control over energy, privacy, and cost.
Why it’s different from AI in the cloud :
Latency: decisions happen near-instantly (no round-trip to a server).
Privacy & compliance: raw data (images, audio, biometrics) can stay on the device.
Reliability: works even with poor or no connectivity.
Efficiency: far less bandwidth and infrastructure cost.
How it works (the edge pipeline)
1. Sense - read data from sensors (camera, IMU, mic, temp, vibration).
2. Preprocess - filter, normalize, or extract compact features.
3. Infer - run a compact ML model (classification, detection, anomaly scoring).
4. Act - drive an actuator, update a UI, trigger an alert, or log locally.
5. (Optional) Learn/Adapt , update thresholds or fine-tune models on-device/federated.
Where it runs
TinyML on microcontrollers (milliwatts):
ARM Cortex-M0/M33 class devices executing quantized models (e.g., keyword spotting, gesture recognition, anomaly detection) via TensorFlow Lite for Microcontrollers.
Edge AI on NPUs/accelerators (watts):
Linux-class boards (e.g., Raspberry Pi 5 AI Kit with Hailo) running heavier vision/speech models (object detection, tracking, segmentation) with dedicated neural processing.
Typical use cases
Industrial: defect detection, predictive maintenance from vibration, safety zones.
Consumer & wearables: wake-words, gesture control, personalized assistants.
Smart spaces: people flow, occupancy, energy optimization, local analytics.
Robotics & mobility: obstacle detection, grasp planning cues, sensor fusion.
Healthcare & well-being: on-device monitoring with privacy by design.
Design considerations
Model efficiency: quantization (int8), pruning, compact architectures (Conv1D/DS-CNN/GRU).
Budgets: RAM/Flash limits, frame rate targets, thermal/power envelopes.
Security: secure boot, signed models, encrypted storage, least-privilege I/O.
Observability: local metrics (latency, throughput, accuracy drift), safe fallbacks.
Updates: over-the-air model/app updates, versioning, rollback plans.
How this fits the Embedac learning path
Start with TinyML: build intuition for constraints by deploying small models on ARM Cortex-M0/M33 boards (gesture/keyword/anomaly).
Scale to NPU Edge AI: move to Raspberry Pi 5 AI Kit with Hailo for real-time vision/speech at higher accuracy and throughput.
Capstone: sense → preprocess → infer → act on-device, with measurable latency, power, and privacy benefits.
Edge AI turns devices from passive data collectors into real-time, privacy-respecting decision-makers from the lab bench to production, from TinyML to NPU-accelerated systems.