Dropout is repurposed as a gradient compression tool to make influence function computation scalable to large models while aiming to retain key influence signals.
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
UNVERDICTED 1representative citing papers
citing papers explorer
-
Toward Efficient Influence Function: Dropout as a Compression Tool
Dropout is repurposed as a gradient compression tool to make influence function computation scalable to large models while aiming to retain key influence signals.