Federal University of Rio Grande do Norte

Technology Center

Department of Computer Engineering and Automation

Embedded AI

References

• :books: Daniel Situnayake and Pete Warden. TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers. [Link]
• :books: Gian Marco Iodice. TinyML Cookbook: Combine Artificial Intelligence and Ultra-low-power Embedded Devices to Make the World Smarter [Link]
• :books: Aurélien Géron. Hands-On Machine Learning with Scikit-Learn, Keras, and TensorFlow [Link]
• :books: François Chollet. Deep Learning with Python [Link]

Lessons

Week 01: Course Outline

• Machine Learning Fundamentals
  • You'll learn how machine learning models work, how to build them, and how to optimize them. By the end, you’ll know the basics behind building models that will make data-driven predictions.
  • :hourglass_flowing_sand: Estimated time: 10h
• Git and Version Control
  • You'll learn how to: a) organize your code using version control, b) resolve conflicts in version control, c) employ Git and Github to collaborate with others.
  • :facepunch: getting a git repository.
  • :hourglass_flowing_sand: Estimated time: 5h
• Complementary materials
  • Google Colab Introduction
  • Google Colab Cont. [optional]
  • :hourglass_flowing_sand: Estimated time: 2h

Week 02: TinyML Fundamentals

• Why our business need AI? And bigger is not always better!!

• How do we enable TinyML?

  • Three fundamental steps to explore a TinyML solution
    • Input
    • Processing
    • Output and final remarks
    • :hourglass_flowing_sand: Estimated time: 30min to 1h.
  • :page_facing_up: Further reading paper
    • Vijay Janapa Reddi et al. Widening Access to Applied Machine Learning with TinyML
    • :hourglass_flowing_sand: Estimated time: 4h

• Machine Learning Fundamentals

  • What is Machine Learning (ML)?
  • ML types
  • Main challenges of ML
    • Variables, pipeline, and controlling chaos
    • Train, dev and test sets
    • Bias vs Variance
  • :hourglass_flowing_sand: Estimated time: 2h

• Calculus For Machine Learning

  • You'll learn how to: a) define mathematical functions using calculus; b) employ intermediate machine learning techniques.
  • :hourglass_flowing_sand: Estimated time: 6h

Week 03: TinyML Challenges

• What are the challenges for TinyML?
• AI lifecycle and ML workflow
  • AI lifecycle introduction
  • AI infrastructure
  • A typical ML workflow
  • A TinyML workflow
  • :hourglass_flowing_sand: Estimated time: 30min
• ML evaluation metrics
  • How to choose an evaluation metric?
  • Threshold metrics
  • Ranking metrics
  • :hourglass_flowing_sand: Estimated time: 1h
• Linear Algebra For Machine Learning
  • You'll learn how to: a) Understand the key ideas to understand linear systems; b) Apply the concepts to machine learning techniques.
  • :hourglass_flowing_sand: Estimated time: 6h
• :page_facing_up: Further reading paper
  • Visal Rajapakse et al. Intelligence at the Extreme Edge: A Survey on Reformable TinyML
  • Sam Leroux et al. TinyMLOps: Operational Challenges for Widespread Edge AI Adoption
  • :hourglass_flowing_sand: Estimated time: 10h

Week 04: Deep Learning Fundamentals I

• The big-picture
• Introduction
  • The perceptron
  • Building Neural Networks
  • Matrix Dimension
  • Applying Neural Networks
  • Training a Neural Networks
  • Backpropagation with Pencil & Paper
  • Learning rate & Batch Size
  • Exponentially Weighted Average
  • Adam, Momentum, RMSProp, Learning Rate Decay
  • :hourglass_flowing_sand: Estimated time: 6h to 8h
• Hands on DL fundamentals
  • You'll learn how to: a) Understand how neural networks are represented; b) understand how adding hidden layers can provide improved model performance; c) Understand how neural networks capture nonlinearity in the data.
  • :hourglass_flowing_sand: Estimated time: 8h

Week 05: Deep Learning Fundamentals II

• A first image classification model using MLOps best practices
• Project :star2: :smiley_cat: :dog: :panda_face:
  • Explore MLOps tools
  • Hyperparameter tuning using Sweeps
  • Compare MLP vs KNN
  • :hourglass_flowing_sand: Estimated time: 8h

Week 06: Convolutional Neural Networks

• Previously on last weeks
• CNN Fundamentals I
  • Convolution with OpenCV and Python
• CNN Fundamentals II
  • Motivation
  • Convolution Layer
  • Convolution Layer - Case Study TinyVGG
  • Pooling Layer
  • Fully-Connected Layer
  • :hourglass_flowing_sand: Estimated time: 2h
• CNN Fundamentals III
  • Batch Normalization Fundamentals
  • Batch Normalization Math Details
  • Batch Normalization - Case Study
  • Dropout
  • :hourglass_flowing_sand: Estimated time: 1h

Week 07: Using CNN to Classify Images

• A MLOPs pipeline using Tensorflow, Keras, Wandb
  • Preprocessing
  • Data segregation
  • Train
  • Test

Week 08: Going Deeper with CNN

• Study of Classical Architectures
• LeNet-5
  • Best practices
  • Extensions using: batch normalization, dropout, data augmentation
  • Sweepy (hyperparameter tuning)

Week 09: Going Deeper with CNN II

• AlexNet
• VGG and Inception

Week 10: Transfer Learning

• Feature extractor and fine-tuning
• Hands on

Week 11: Edge Impulse crash course

• A brief overview of Edge Impulse Platform
• Data Acquisition
• Create a impulse design and a preprocessing task
• Training
• Understanding training evaluation metrics
• Model testing
• Live classification using a mobile phone
• AutoML configuration using EON Tuner
• Understanding the results of EON Tuner and versioning the model
• Set a primary model using EON Tuner and Transfer Learning
• Training an EON Tuner primary model using transfer learning
• Final remarks

Week 12: TFLite Optimizations and Quantization

• Post Training Quantization (PTQ)
• Introduction to TensorFlow-Lite
• PTQ of MNIST
• A regression model using TensorFlow-Lite
• Case study using Wandb developed by Ishan Dutta et al.
  • Optimizing Models with Post-Training Quantization in Keras - Part I
  • Optimizing Models with Quantization-Aware Training in Keras - Part II