[spam][crazy][ot] MCBoss Spinoffs 3

[Karl Semich]

— -
ok so i saw a realworld person with handmade feed attached to
amputations right below the knees and i got a rockawesome idea for a
puzzle: free homebrew robotic prosthetic limbs that respond to user
intention accurately.

i think i’ve solved big chunks of the puzzle in outline form. the core
unique part of interest is responding to user intention.

1. assume a shared community pretrained model. for initial users, this
is a blank model.
2. prompt the user to try to move their limb while indicating what
they are trying to do.
3. size the model for this data only and train it maximally.

4. now the user can use the poorly-trained limb so long as conditions
precisely match the training conditions. get them using it, and
collect more data as they do.
5. augment the data taking the times the prosthetic is functioning
effectively and use its own outputs as labels, but mask the inputs it
is using to judge them. increase the size of the model for the amount
of newly available data, and continue training online.

the point of 5 is to make the model highly robust in new conditions.
similar to dropout during pretraining, it forces it to learn many
different indicators of what is correct.

doing the training online hopefully helps the user learn ways to
control the prosthesis at the same time as it is learning to infer
what they want. [some further consideration could engage here to find
other ways to do this]

for further effectiveness, since so much training is going on, a
metamodel can be trained to make this adaptation more rapid.

not all idea reached concept writing but what did reach seems
reasonable to start with. concerns of interest are what are the
channels of information and what is the actuation that responds to the
user. conveniently, these can be pretty much anything.