## TL;DR
AI for IoT brings intelligence to the physical world — from smart thermostats learning your preferences to industrial sensors predicting equipment failure. Federated learning enables AI training across millions of devices without centralizing data, while TinyML compresses models to run on microcontrollers smaller than a grain of rice.

## Core Explanation
IoT architecture: sensors (temperature, motion, camera, microphone) → edge devices (microcontrollers, gateways) → cloud. AI deployment options: (1) Cloud AI — send all data to cloud, run large models (highest accuracy, highest latency, privacy risk); (2) Edge AI — run inference on edge gateway (Raspberry Pi, Jetson Nano); (3) On-device AI (TinyML) — run inference on the sensor's microcontroller itself (ARM Cortex-M, <1MB RAM, <1mW power). Federated learning: instead of sending raw sensor data to a central server (privacy violation for home/health data), each device trains a local model on its own data and sends only model updates (gradients/weights) to a central server that aggregates them (FedAvg algorithm). The aggregated global model improves from all devices' data without ever seeing the data itself.

## Detailed Analysis
TinyML pipeline: (1) Train large model (TF/PyTorch); (2) Quantize — reduce 32-bit float weights to 8-bit integers (post-training quantization, quantization-aware training); (3) Prune — remove unimportant weights; (4) Knowledge distillation — train a small "student" model to mimic a large "teacher"; (5) Convert to microcontroller format (TensorFlow Lite Micro, ONNX Runtime). Result: 50-500KB models running at 5-50ms inference time. ScienceDirect 2025 FL+TinyML survey: combining FL (privacy) + TinyML (efficiency) enables intelligent IoT without privacy-efficiency tradeoffs. Nature Comm 2025 personalized IoT: federated meta-learning (Per-FedAvg, pFedMe) learns a good initialization that adapts quickly — Model-Agnostic Meta-Learning (MAML) applied in federated setting, enabling 5-shot personalization. MDPI 2025 FL-for-IoT survey documents 15+ FL frameworks for IoT (TensorFlow Federated, PySyft, Flower, FedML). Applications: (1) Smart home — learning occupancy patterns for energy optimization without sending home data to cloud; (2) Wearables — personalized health monitoring (arrhythmia detection, fall detection) adapting to individual baselines; (3) Industrial IoT — predictive maintenance across factory machines; (4) Agriculture — soil and crop monitoring sensors. Key challenges: heterogeneous devices (phone vs. microcontroller), communication efficiency (sending model updates over low-bandwidth IoT protocols like LoRaWAN), and Byzantine robustness (malicious devices poisoning the global model).

## Further Reading
- TensorFlow Lite for Microcontrollers (TinyML)
- Flower: A Friendly Federated Learning Framework
- Edge Impulse: TinyML Development Platform