How To Implement Efficient Sharpness Aware Minimization

– () , . , , – ./

/

, /
% /
, , /
, /
(ρ) — . . /
/

– /

– “//.//.” “” “” . ()/ . , , ./

ε- . , ./

() . “//..////-.” “” “” / ./

– /

. , ./

. “//..//-” “” “” / ./

, — . ./

– /

. , θ ε ρ · / ||||, ρ . , θ + ε θ ./

/

/ θ/ ||ε||≤ρ/ (θ + ε)/

//

/ ε̂() ρ · ∇() / ||∇()||/

/ ← – η · ∇( + ε̂())/

. ρ ./

/

. , , — . , ./

. “//.///” “” “” / – . , “//./////—” “” “”‘ / – ./

. ., (ρ) . ., . ./

/ /

. . ./

ρ . , . ./

, . ./

/

. () ./

– () . ./

, – – . , – – ./

/

. ρ ./

– . ./

. ./

/

(ρ) /

ρ . . , . – ρ ./

/

, – () . ./

/

-% . ./

– – /

– , -% . – ./

/

, . , ./

/

. ./

/

. , ρ . ./

– /

“//./////—” “” “” / “//.///” “” “” /. ./