What would be the best ML model an Arduino nano 33 ble sense can handle?

Below is the information I get from Havard TinyML course.

We suggest you leave the MODEL_ARCHITECTURE as tiny_conv the first time but if you would like to do this again and explore additional models some options are: single_fc, conv, low_latency_conv, low_latency_svdf, tiny_embedding_conv.

Assume I want to train a KWS model with 4 custom keywords, beside that, the arduino will send the inference result via BLE. I have tried tiny_conv and it can run with no issue. However, what would be the best model inside it’s capability for this application?

The “Best” model will be very application dependent. The exact size and number of layers is almost impossible to know before you start trying to train and then test models. There is a whole research area of AutoML (AutoML | AutoML) that is trying to figure this out!

That said, in general for image processing convolutional networks seem to work the best. And similarly for other application domains there are certain kinds of models that often work well. Hope that helps!

1 Like