Aug 2026
Fisher8: Stabilizing Neural Heteroscedastic Regression via Output-Layer Fisher Geometry
Fisher8 is derived, an output-layer gradient correction that reorients and rescales updates using Fisher geometry rather than Euclidean geometry, and it is shown that prior stabilizers converge on overlapping components of this geometric correction.
Sumedh Vemuganti, Nickvash Kani
· arXiv.org · 0 citations