Hey there guys. Looks like you all asked some great questions already! However, I was still hoping to get some clarification on a thing or two.
At 15:24:
@vclay says that the LM suggests a location relative to the body, then the motor system translates that, driving movement. My question is this: is it the motor unit itself performing the end effector → target position transforms, or is it being given this data by the LM?
At 1:14:28:
You guys start talking about sensor displacement awareness here. Somebody says “Monty obviously knows the displacements”. I was hoping for some clarity on some things with this statement…
When we say “Monty” are we talking about the totality of all sensors (both SM & LM)? If so, then when we say “Monty knows the sensor displacements for all sensor modules,” are we really saying that an all-to-all call function returns us these data?
…I was thinking back to one of your more recent videos (the one concerning voting mechanisms and optimizations), and it got me wondering…
If by “Monty” we mean “the totality of the system” then wouldn’t it make more sense for us to view something like Column ID retrieval hierarchically, as opposed flatly organized (ergo the all-to-all call)? For example, if we view neurological columns in the brain (say, in v2 or v4) wouldn’t those columns understand the relational connections between lower level columns found in V1? Couldn’t those v2/v4 columns return the relevant IDs to those V1 columns, removing the computational cost associated with an all-to-all comparison?
Edit: I guess my question for that second part is: wouldn’t a lot of your guys’ optimization issues be solved by applying hierarchical constructions on top of your baseline approach to voting (that being all-to-all)?
So for instance, when you first initialize a Monty system and run through an episode, the LMs may need to resort to all-to-all call functions to begin with. But as your continue through your epoch, and move into new ones, not only should the LMs begin to remember their historical neighbors, by way of local caching (see comment here), but there should be higher-order LMs above them which understand the relationships of entire functional groups, no?