TinyML: Production model development (IDA København)

TinyML: Production model development (IDA København)
In this third workshop on TinyML, we go one step further with the audio event detection model and explore techniques for improving the model all the way to production quality. In particular, we will learn how to properly evaluate a model’s performance ...

In this third workshop on TinyML, we go one step further with the audio event detection model and explore techniques for improving the model all the way to production quality. In particular, we will learn how to properly evaluate a model’s performance and how to achieve the highest possible accuracy while minimizing the latency and resource usage on a tiny device.

While standard ML workflows can get you quite far, optimizing your machine learning model to become production-ready requires some additional insight and effort. In this workshop, we will focus on three important aspects of model improvement:

  • Performance evaluation. For this part, we will go through the different performance metrics, both for accuracy and resource usage and how to use them to pinpoint where there is room for improvement.
  • Accuracy improvement. In this part, we will learn about the typical causes of ML model misclassifications and the techniques for maximizing accuracy. In particular, we will focus on improving the data and labels. These techniques include:
    • Advanced feature extraction
    • Iterative data set improvement
    • Data augmentation
    • ML-assisted labeling
  • Resource usage optimization. For the final part, we will ensure that your model is as fast as possible and uses as little power and memory as possible. We will briefly look at how different compilers can optimize your model automatically by quantization, pruning and utilizing the microcontroller’s hardware acceleration cores. However, there are also some manual steps you can take to significantly optimize your model, including:
    • Finding the optimal model architecture
    • Hyper-parameter tuning
    • Accuracy/size trade-off
    • Streaming stateful models

In the practical part of the workshop, we will apply these techniques to our running audio event detection use case, but again you are welcome to work with your own case, and we will be happy to assist you the best we can.

This workshop will run in AU Herning in parallel via a video link

All workshops and dates
This is the third of 3 workshops, that is also organised and streamed to AU Herning. You can sign up for one of all three workshops here:

November 20 - Introduction to TinyML
Workshop 1 - IDA København
Workshop 1 - AU Herning

December 4 - TinyML Data preparation, training and deployment
Workshop 2 - IDA København
Workshop 2 - AU Herning

December 18 - TinyML Production model development
Workshop 3 - IDA København
Workshop 3 - AU Herning

ANYONE CAN JOIN - Most of IDA's events are open to everyone, but you must have an IDA user profile in order to participate. It's free and non-binding - create a user profile now. As a member of IDA, you get a discount on many of our events, so please sign in with your member login when you sign up for an event. Read more about the benefits of an IDA membership here.

Information
  • When

    18. dec. 2023 17:30 - 21:00
  • Where

    IDA Conference, Kalvebod Brygge 31-33, 1780 Copenhagen V

  • Registration Deadline

    17. dec. 2023 - 23:55

  • Organizer

    IDA Embedded

  • Available Seats

    5

  • Event Number

    352211

Price
  • Vacant

    0 kr.

  • Member

    0 kr.

  • Senior member

    0 kr.

  • Student member

    0 kr.

  • Participant, not a member of IDA

    300 kr.