Good webinar, thanks for presenting more on this topic.  I appreciated all the time you spent answering our questions.  Because time was so you did go through the demo quickly.  I watched this tinyml video from Jan at Edge Impulse, it looks a lot like the demo you were showing today.  The video takes us through recognition training, and was especially helpful to me because he then went into anomaly detection. Very nice.

https://www.youtube.com/watch?v=FseGCn-oBA0&t=494s

Success!  My nano-33-BLE-sense has collected motion data to edgeimpulse.com.  The instructions are excellent.  I'm using the arduino nano 33 BLE sense, so I followed the instructions at

https://docs.edgeimpulse.com/docs/arduino-nano-33-ble-sense

But, my fault, I tried it first with my Windows7 laptop, and it became clear Win7 is too old, I got stuck halfway.  Too many errors from the install routines.

So, on to my raspberry Pi4 instead.  That worked much better.  But I did have a couple problems that stumped me for a few minutes.  The downloaded arduino-cli had to be copied to /usr/local/bin (or to somewhere in my $PATH).

And second thing, probably due to user error, but, the instruction say to run:

     ~ $ edge-impulse-daemon

does not run the daemon.  I found edge-impulse-daemon in ~/.npm-global, and then ran it with:

  ~/.npm-global $ ./edge-impulse-daemon

and all was good.  Now on to record motion data and train my model.

Really enjoying the edge impulse interface for learning to use my nano BLE sense for AI.  The visual 3D model of the separation of the labeled data is nice, you get a real feel for the data you've recorded.

Reviewing Jan's video was very helpful, the second (and third) time to pick up details I missed.  The motion classification example didn't work my first time, but now it's almost perfect.  The first time through I had (incorrectly) assumed that I had to compensate the acceleration data for the earth's gravity, by training the system at "idle" in several different orientation.  That lead to the result having no separation in the model.  Confusion in the "confusion matrix".  Review video: Oh, they already knew to take earth's gravity into account.  Delete the multiple orientations, and now it works great.

Now, on to my own classification to detect anomalies.

Really enjoying the edge impulse interface for learning to use my nano BLE sense for AI.  The visual 3D model of the separation of the labeled data is nice, you get a real feel for the data you've recorded.

Reviewing Jan's video was very helpful, the second (and third) time to pick up details I missed.  The motion classification example didn't work my first time, but now it's almost perfect.  The first time through I had (incorrectly) assumed that I had to compensate the acceleration data for the earth's gravity, by training the system at "idle" in several different orientation.  That lead to the result having no separation in the model.  Confusion in the "confusion matrix".  Review video: Oh, they already knew to take earth's gravity into account.  Delete the multiple orientations, and now it works great.

Now, on to my own classification to detect anomalies.

I also tried Edgeimplse on Arduino nano 33 BLE sense. This is  a good place for someone who is new to ML and Deep learning. Best thing is we can see the data. Split the data and manage the data in an effective manner.

I tried voice recognition. It didn't work but the reason was that I did not have enough data. With more data, I am sure it will work nicely.

BTW, Do you have any idea weather we can combine conventional C/C++ code for sensors with the NN model and upload the binary to the board? So that we can get the NN model running at the same time sensors are displaying values?

Ofcourse you can. Just export the model as a cpp library and use it as you would any other library.

Regards,

Moinak