From random guesses to feedback-informed weightings: we learn to crumble the cookie (so many readings!) back into correctly labeled movements (archetypal gestures). The cookie uncrumbles.
The distance between "bullseye and arrow hit point" generalizes to all crumbs (crumbs in all dimensions) as does the computation to minimize this distance (e.g. "gradient descent" -- one of the methods).
The crumbs look more and more like the original cookie, as the weightings get shifted, epoch after epoch on screamingly fast chips. We need already judged and labeled originals so the feedback keeps happening. Supervised learning continues.
From randomly guessing focus group committees (new enrollees) to finely refined teams with proved track records (weighted and ranked accordingly), do we mitigate this metaphysical distance, twixt our predictions and what is.
I picture a whole skyscraper of committees, as floors of (levels of) "perceptron neural nodes" each signaling to the next floor above it and, and so on (deep neural net), with a broadcast tower atop the building (some KPOJ) sharing its newly synthesized view of the world (its brand of enlightenment programming, its model).
[ with thanks to Nathaniel Bobbit ]
No comments:
Post a Comment