AI Roadmaps
← All roadmaps
AdvancedRoadmap 24 · 14 steps · 5h 27m of video

TinyML & Edge AI

Shrink models and run them on microcontrollers, phones, and small boards: pruning, quantization, distillation, TFLite Micro, Edge Impulse, ONNX, and ExecuTorch.

Before this: Deep Learning (training a CNN in PyTorch or Keras). Hardware steps use a microcontroller board such as an ESP32-CAM or Arduino Nano 33 BLE Sense, a Jetson, or a Raspberry Pi; the rest run on a laptop.

  1. 6 min
    Step 1

    What TinyML is

    Why running ML on tiny, battery-powered chips matters.

    What is TinyML?

    Arm®

  2. 1h 03m
    Step 2

    Why efficient ML

    MIT's course intro: the gap between model size and hardware.

    EfficientML.ai Lecture 1 - Introduction (MIT 6.5940 Fall 2026)

    MIT HAN Lab

  3. 42 min
    Step 3

    Pruning and sparsity

    Remove weights a model does not need, and keep accuracy.

    EfficientML.ai Lecture 3 - Pruning and Sparsity (Part I) (MIT 6.5940 Fall 2026)

    MIT HAN Lab

  4. 23 min
    Step 4

    int8 quantization

    Store weights and activations in 8 bits instead of 32.

    Understanding int8 neural network quantization

    Oscar Savolainen

  5. 6 min
    Step 5

    Knowledge distillation

    Train a small student model to copy a big teacher.

    Knowledge Distillation - Explained

    DataMListic

  6. 1h 03m
    Step 6

    TinyML on microcontrollers

    MCUNet fits vision models into kilobytes of memory.

    EfficientML.ai Lecture 10 - MCUNet and TinyML (MIT 6.5940 Fall 2026)

    MIT HAN Lab

    mit-han-lab/tinyengine★ 969
  7. 13 min
    Step 7

    Optimize TensorFlow Lite models

    Post-training quantization and pruning in the TFLite toolchain.

    Optimize your TensorFlow Lite models | Session

    TensorFlow

  8. 19 min
    Step 8

    TensorFlow Lite for Microcontrollers

    Deploy a first model to a microcontroller, end to end.

    TinyML: Getting Started with TensorFlow Lite for Microcontrollers | Digi-Key Electronics

    DigiKey

    tensorflow/tflite-micro★ 3.2k
  9. 28 min
    Step 9

    Keyword spotting in 30 KB of RAM

    Collect audio, train, and deploy a wake-word model.

    Build Your Own ML-Powered Keyword Spotting Model in 30KB RAM

    Edge Impulse

  10. 17 min
    Step 10

    Object detection on an ESP32-CAM

    Run a tiny detector on a cheap camera board.

    How to do Object Detection using ESP32-CAM and Edge Impulse YOLO Model

    Circuit Digest

  11. 8 min
    Step 11

    ONNX Runtime for generative models

    See how ONNX Runtime GenAI runs language models on a device.

    Introduction to ONNX Runtime

    NVIDIA Developer

  12. 25 min
    Step 12

    On-device PyTorch with ExecuTorch

    PyTorch's runtime for phones and embedded chips.

    ExecuTorch 1.0: General Availability Status for Mobile and Embedded...- Mergen Nachin & Cemal Bilgin

    PyTorch Foundation

    pytorch/executorch★ 5.2k
  13. 5 min
    Step 13

    Get started with a Jetson Orin Nano

    Set up a small GPU board for edge vision and LLMs.

    Getting Started with the Jetson Orin Nano Developer Kit

    NVIDIA Developer

  14. 9 min
    Step 14

    Run an LLM on a Raspberry Pi 5

    What a small language model can do on a cheap single-board computer.

    Gemma 4 on Raspberry Pi 5: A Surprisingly Usable Local AI Setup

    Zero to MVP

    ggml-org/llama.cpp★ 135.8k

Build it

Watching isn't learning. Pick one and put it on GitHub.

Browse all roadmaps →