Machine learning, edge computing and 32 bits

It was a pleasure to take part in the ANNPR conference together with ScorePad. ANNPR stands for Artificial Neural Networks in Pattern Recognition. Thanks again to the colleagues at our long-standing partner ZHAW for organising it so well; working with you is a pleasure.

For us, with our embedded focus, it gets interesting as the overlap grows between microcontrollers (32 bit at 240 MHz or less, say) and machine learning algorithms. What the controller does varies a great deal with the task and the setting. Sometimes it is no more than a sensor interface to a cloud database, which for us means collecting data, checking it for plausibility, encrypting it and sending it over the protocol of choice. The rest is then usually handled by web services or apps.

Edge computing is moving the other way, more and more into the foreground: the collected data is pre-processed on the device and only then sent to the server for final classification. A fine example is Google's Edge TPU, which grew out of the offline speech recognition in the Google Assistant on Android devices and now makes machine learning tasks feasible at the lower end of embedded performance as well.

Do you have an embedded task close to machine learning? Talk to us, we look forward to it.

Wishing you plenty of lateral thinking, your querdenker engineering.