IMPACT derives a closed-form low-rank activation reconstruction from an importance-weighted covariance matrix to achieve higher compression ratios than standard methods while maintaining model accuracy.
GroupReduce: Block-Wise Low-Rank Approximation for Neural Language Model Shrinking
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
-
IMPACT: Importance-Aware Activation Space Reconstruction
IMPACT derives a closed-form low-rank activation reconstruction from an importance-weighted covariance matrix to achieve higher compression ratios than standard methods while maintaining model accuracy.