How to implement Machine Learning using nRF51822 and Accelerometer ADXL345

RE: How to implement Machine Learning using nRF51822 and Accelerometer ADXL345

mantap — Sun, 01 Apr 2018 23:37:25 GMT

thanks for answering, but If you can give me example code to implement this machine learning to my module. it's very usefull..

because I still confuse to separate from main program that recognize some pattern calibrate function program.

please help...

RE: How to implement Machine Learning using nRF51822 and Accelerometer ADXL345

mantap — Sun, 01 Apr 2018 23:34:52 GMT

thanks for your answering, can you give me the example code for mbed compiler, that using statistic machine learning such as K-nearest neighbor, naive bias, or what ever that possibly to make it happen..

If I using python from pc for collect data, get learning the classify pattern such as standing, walking, and climbing up. then get recognize it to my module..
what code I suppose to try?..

RE: How to implement Machine Learning using nRF51822 and Accelerometer ADXL345

Turbo J — Sun, 01 Apr 2018 10:19:01 GMT

As you noticed, the NRF51 does not have enough "beef" to do expensive AI calculations by itself.

You have now 2 options:

Transfer the data to something bigger like a PC or mobile and do the AI there.
Ditch the AI and use a much less expensive classic statistics algorithm.

RE: How to implement Machine Learning using nRF51822 and Accelerometer ADXL345

Janek Mann — Sun, 01 Apr 2018 08:32:13 GMT

Machine learning for resource-constrained environments is a rather complicated topic. You can't just expect to take an algorithm developed in Tensorflow and port it across to a Cortex-M0.

You could implement a model of the resource constraints on the NRF51822 and attempt to train a model within those constraints, but I am not aware of any out-of-the-box solutions for this.

For human activity recognition there is a decent body of existing research on hand-designed implementations that can run in a microcontroller.

Alternatively, a popular approach is to implement low-power feature extraction on the MCU and use a model in say a phone app to perform the actual recognition. But even if you want to implement the recognition on the MCU the separation of feature extraction and recognition is a useful approach in a resource-constrained environment.

Hope that gives you some useful terms to Google, and don't forget about scholar.google.com too!