Statistical Learning
Rethinking gradient sparsification as total error minimization
We identify that the total error -- the sum of the compression errors for all iterations -- encapsulates sparsification throughout training.
4: fort {1,2,,E}do 5: StepP(t) P(t 1)+αt P(t 1) h ℓattack f A+P(t 1), X; θ =train(A+P(t 1),X,y),y i
As introduced in 2, attacks may perturb the adjacency matrixA, the feature matrixX, or both.
HowDataAugmentationaffectsOptimizationfor LinearRegression
Our approach interprets augmented (S)GD as a stochastic optimization method foratime-varying sequence ofproxylosses.
TopP&R: Robust Support Estimation Approach for Evaluating Fidelity and Diversity in Generative Models
However, the reliability of these estimates has been overlooked, even though the quality of the evaluation hinges entirely on their accuracy.
