-
find out if there's a better way to get the activations from Axon
-
set up different Supervisor trees for different targets (in particular, create dummy Knob and Display servers for local dev - display could even do a crude web mockup of the thing)
-
set up dense layers to clamp bias to 0 (or at least "fake" the activations by using the layer norm outputs rather than the raw dense_1 outputs)
-
(maybe) add "wiggle back and forth" detection for reset training?
-
(if necessary) add debounce logic to the rotatry encoder
-
rename Knob to Input (in anticipation of having multiple inputs)
-
write up a blog post about the design & build