varchanaiyer / study-group

Syllabus, schedule, resources etc. for TinyML study group

TinyML Study Group

Syllabus

• Day 1: Introduction to TinyML (13th Feb, 2021. 9am SGT)

  • What is TinyML?
  • Why TinyML is important?
  • TinyML Hardware
    • Microcontrollers
    • Accelerators: Neural Compute Stick, EdgeTPU
    • Mobile
  • TinyML Algorithms
    • Quantization, Knowledge Distillation, Pruning, Early Exits
  • TinyML Software
    • TFLite
    • ONNX
    • OpenVINO
  • Why learn TinyML?
    • Job prospects, research, projects, community
  • General information about this study group
  • Homework: Find an algorithm/hardware/framework not mentioned today

• Day 2: Basics of Running Neural Networks at the Edge (27th Feb, 2021. 9am SGT)

  • Edge Performance Metrics
    • Latency
    • Throughput
    • Size
    • FLOPs
    • Other Metrics
  • Measuring performance metrics
  • Which metrics to optimize your network for?
  • Homework: Read a paper that compares performance of different networks, hardwares, frameworks

• Day 3: Quantization (13th March, 2021. 9am SGT)

  • What is Quantization?
  • Different Quantization Specs
    • Weight Quantization
    • Weight and Activation Quantization
  • How does Quantization improve performance at the edge?
  • Quantization using TFLite
  • Running a Quantized network on a hardware
  • Measuring Quantization Performance
    • How much does quantization improve performance?
  • Homework: Quantize a network using OpenVINO and measure its performance

-Day 4: Edge Hardware (27th March, 2021. 9am SGT)

• Introduction to the Intel Neural Compute Stick 2 (NCS)

• NCS Internals: How does it work?

• How to use OpenVINO to interface with the NCS

• Using Multiple NCS devices

• Homework: Compare the Latency and Throughput of NCS, CPU and GPU

• Day 5: Pruning (10th April, 2021. 9am SGT)

  • What is Pruning?
  • Different Pruning Methods
  • How does Pruning improve performance at the edge?
  • Pruning using TFLite
  • Running a pruned network on a hardware
  • Measuring Pruning Performance
    • How much does Pruning improve performance?
  • Homework: Prune a network and measure its performance

• Day 6: Research Day - Early Exits in Neural Networks (24th April, 2021. 9am SGT)

  • Read and discuss a paper on Early Exits

• Day 7: Edge Hardware: EdgeTPU and EdgeTPU Accelerator (8th May, 2021. 9am SGT)

  • Introduction to the EdgeTPU
  • EdgeTPU Internals: How does it work?
  • How to use TFLite to interface with the NCS
  • Homework: Compare the Latency and Throughput of EdgeTPU and the EdgeTPU accelerator

• Day 8: Project 1 - Train and Deploy a model at the edge (22nd May, 2021. 9am SGT)

• Day 9: Security and Privacy in Edge Computing

  • Security of Hardware
  • Security of Data
  • Security of Networks running on Edge Deployed Hardware
    • Cold Boot Attacks
    • Side Channel Attacks

• Day 10: Edge Hardware: TinyML on an Arduino

  • Introduction to the Arduino
  • Arduino Internals: How does it work?
  • TinyML on Arduino: link