For a two-layer CNN with a signal-noise data model, FedAvg test error depends on the number of filters misaligned at initialization, and pre-training helps mainly by reducing that number.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Initialization Matters: Unraveling the Impact of Pre-Training on Federated Learning
For a two-layer CNN with a signal-noise data model, FedAvg test error depends on the number of filters misaligned at initialization, and pre-training helps mainly by reducing that number.