A learning-rate-aware matrix factorization for correlated noise in private SGD yields better accuracy than prefix-sum factorizations on CIFAR-10 and IMDB while remaining memory-efficient.
Antti Koskela, Joonas J ¨alk¨o, Lukas Prediger, and Antti Honkela
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Learning Rate Scheduling with Matrix Factorization for Private Training
A learning-rate-aware matrix factorization for correlated noise in private SGD yields better accuracy than prefix-sum factorizations on CIFAR-10 and IMDB while remaining memory-efficient.