Test if we should reduce learning rate

[Karlo-Emilo]

From micro:bit: "We have reduced the learning rate and that’s resolved an issue we saw when going above 5 classes where models got rapidly worse. This has given us some really great results with ~8 data samples for each class"

[r59q]

The learning rate is adjustable. I believe we can easily reduce the default learning rate, given how quickly the value converges.

Which version should these apply to, is a new feature? I.e features.json value

[r59q]

Make 0.1 the default learning rate