AI@EDGE

# Usergroup meeting 3 AI@EDGE

Date: 29 June 2021
Location: Online - Teams

  • Present:
    • Kris Bellemans (E.D.&A.)
    • Nicolas Maes (Picanol)
    • Peter Papics (Transport & Mobility Leuven)
    • Luc Buydens (Melexis)
    • Frank Allemeersch (Sensotec)
    • Chandu Kancharla (KU Leuven - Brugge)
  • Excused:
    • Fabrice Verhaert (6Wolves)
    • Gert Van de Wouwer (Digipolis)
    • Dieter Therssen (DSP Valley)
    • Patrick Smout (E.D.&A.)
    • Patrick Puissant (Picanol)
    • Kris Vanherle (Transport & Mobility Leuven)
    • Ben Minnaert (Odisee)
    • Nick Destrycker (Edgise)
    • Willem Romanus (6Wolves)
    • Edouard Charvet (Scioteq)
    • Anke Van Campen (VLAIO - project advisor)
  • Project team members:
    • Toon Goedemé (KU Leuven - De Nayer)
    • Kristof Van Beeck (KU Leuven - De Nayer)
    • Maarten Vandersteegen (KU Leuven - De Nayer)
    • Lotte Hendrickx (KU Leuven - De Nayer)
    • Piet Cordemans (VIVES)
    • Sille Van Landschoot (VIVES)
    • Jonas Lannoo (VIVES)

# Presentation Slides

Preview:

Open the presentation in a new page:
User group meeting slides

# Introduction

A short introduction was given on the status of the project, the progress of the work packages and the work that yet has to be done.

  • Comments:
    • No specific remarks from the user group.

# Frameworks & Edge Impulse

In this chapter the Edge Impulse framework was introduced together with its workflow. The possible options for each step were shown. The supported libraries and methods to deploy your trained model to an embedded device has been explained. Additionally, the supported optimalisation techniques embedded in Edge Impulse were explained, such as quantisation. Edge Impulse also has an optimised compiler which was also explained.

  • Comments:
    • Luc Buydens (Melexis): How long does it take to port a network to embedded hardware? It depends, ranging from a few hours to weeks. Does the framework include all required layers? What’s the knowledge of the engineer? Do we need to create specific functions ourselves?
    • Chandu (KU Leuven): Did you start with a pre-trained network? What if the network changes?
    • Chandu (KU Leuven): What’s the impact on the accuracy when quantizing the network?
    • Chandu (KU Leuven): Were the frameworks chosen specific for this application for speed or possibility for optimization?
    • General: Can we import all networks in Edge Impulse?

# Academic use-case

The academic use case AB writing was fully explained using the workflow in Edge Impulse. Several methods on how to improve your data is also explained, such as using the frequency domain or spectrogram or other relevant parameters that can be derived from your original data. Additionally the influence of number of neurons and model size was shown.

  • Comments:
    • Piet Cordemans (VIVES): If you increase the number of neurons, possibly overfitting might occur. Your graphs do not show this. Should this be investigated?
      • Current network does not reach the overfitting point.
    • Maarten Vandersteegen (KU Leuven): You used a single hidden layer. Did you try multiple layers?
      • Yes, but they did not yield better results.
      • User more layers might also be able to reach the overfitting point.
    • Chandu (KU Leuven): Are there consistent misclassifications? Is there a trend in the errors?
      • The speed of writing, moving the pen before/after writing, ...
    • Chandu (KU Leuven): Was there a particular orientation when holding the pen?
      • Microcontroller up, but position was changed during measurement. Extra testing necessary to see the effect.
    • Toon Goedemé (KU Leuven): When going from 32-bit to 8-bit (factor 4) weight quantization, the memory and flash usage decreases with factor 3.
      • Ok, same magnitude of decrement.
    • Maarten Vandersteegen (KU Leuven): Does the flash usage only includes the model?
    • Piet Cordemans (VIVES): Does this latency take into account the pre-processing time?
      • No.
    • Toon Goedemé (KU Leuven): Why split up A and B in left and right handed. Why not go for a single class?
      • Both are valid options. When starting the project I started this way, I could combine but have chosen for the 5 classes instead of 3. Easy option to combine the classes.
    • Luc Buydens (Melexis): You only need the lower frequencies. To make calculation complexity lower, why don't you cut off the higher frequencies?
      • No options in the framework. Solution: there’s a whole documentation on how to customize and implement your own functions.
      • 1D convolution performed better than 2D, since there is more information.
    • Luc Buydens (Melexis): What information of the spectrogram does the convolution use?
      • 1D convolution on the 2D data will apply the filter over a single axis, after investigating only over the time axis.
      • 2d convolution performs it on both axes simultaneously.

# E.D.&A. use-case

Sille explained the problem statement of the use case from E.D.&A. about the induction cooker touch button. The goals and possible approach was exploited.

  • Comments:
    • No specific remarks from the user group.

# TML use-case

Lotte showed her progress on the TML use case to detect cars, bikes and humans crossing the street. She showed that she can detect and track the objects. Further work has to be done to generalise the network, she needs more data and retrain the model. Also further optimisation of the model is possible.

  • Comments:
    • Peter Papics (TML): In the meantime we have collected a lot more video material (for other development purposes). We would be very happy to share those, so you could have more data to improve the accuracy.
      • Currently the generalization is not that good. For real-life applications
    • Toon Goedemé (KU Leuven): Is the data annotated?
      • Peter Papics (TML): We are working with partners to annotate, need to check this out.
    • Luc Buydens (Melexis): How do you handle the GDPR for this?
      • Peter Papics (TML) Have asked from the relevant commission: there is no issue because the images are not recorded. The telraam devices do not transmit any image data. All processing is happening on the devices themselves, you can not recognize people (face, number plates,...)
      • The videos for training are not shared and only used internally for development.

# Melexis use-case

Maarten gave an update on the status of the Melexis use case. He needs more data and more camera’s will be distributed to capture images.

  • Comments:
    • No specific remarks from the user group.

The meeting was concluded with some advertisement of the workshop that is coming up in December. Networking was done in a virtual space.

← UG Meeting 4 →