edge-ai

TinyML, neural network quantization (INT8), pruning, edge accelerators (Google Coral, Hailo, Jetson), and embedded inference

SKILL.md

GitHub

TinyML & Edge AI

Scope

Edge AI deploys optimized deep learning inference models directly onto resource-constrained microcontrollers and edge accelerators with milliwatt power budgets.

Model Optimization & Quantization

  • Quantization: Converting 32-bit floating point weights ($FP32$) to 8-bit integers ($INT8$): $$q = \operatorname{round}\left(\frac{x}{S}\right) + Z$$ where $S$ is scale factor and $Z$ is zero-point; reduces memory footprint by $75%$ with minimal accuracy degradation.
  • Model Pruning: Removing redundant weights and channels (structured/unstructured pruning).
  • Embedded Inference Engines: TensorFlow Lite for Microcontrollers (TFLM), Edge Impulse, CMSIS-NN, ONNX Runtime.

Tools & Standards

  • Hardware: ARM Cortex-M55/M85 (with Helium vector extensions), ESP32-S3, Google Coral TPU, NVIDIA Jetson.
  • Canonical References: Warden & Situnayake — TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers.