Centrala begrepp
Perceptually-aligned gradients in robust models are explained through off-manifold robustness.
Sammanfattning
Robust computer vision models exhibit perceptually-aligned gradients (PAGs).
PAGs enable generative capabilities like image generation and denoising.
Off-manifold robustness explains PAGs aligning with human perception.
Bayes optimal models demonstrate off-manifold robustness.
Different regimes of robustness impact perceptual alignment and model accuracy.
Various phenomena related to PAGs are identified and discussed.
Theoretical connections between Bayes optimal predictors and perceptual alignment are established.
Empirical verification of off-manifold robustness in robust models is provided.
Signal-distractor decomposition is used to understand the manifold structure of PAGs.
Experimental evaluation confirms theoretical analyses and hypotheses.
Statistik
모델의 입력 그래디언트는 데이터 매니폴드에 거의 일치하는 경향이 있습니다.
Bayes 최적 모델의 그래디언트는 신호 매니폴드에 위치합니다.
로버스트 모델은 신호 매니폴드에 상대적으로 강한 저항력을 가지고 있습니다.
Citat
"Perceptually-aligned gradients in robust models are explained through off-manifold robustness."