7 upvotes
TWIG compresses neural nets with tree-structured discrete weight paths
What is Every Weight Is a Path of Shared Discrete Decisions?
TWIG encodes each weight as a short sequence of binary decisions, a sign, a power-of-two shift, an optional zero gate, and an index into a shared table. The tree-like format lets many weights share upper bits, cutting storage and enabling integer-only arithmetic.
Who it is for
Model developers targeting smaller footprints or faster edge inference can try TWIG to explore new compression-speed trade-offs.