Wanda
A post-training pruning method that requires no retraining. It determines which parameters to remove by jointly considering weight magnitudes and statistical information from input activations, maintaining competitive overall performance at high compression ratios. It is commonly used in large model compression research.
Timeline (last 90 days)
After applying 50% Wanda pruning to Whisper-large-v3, the absolute word error rate gap between the worst-served and best-served groups — specifically between Black/African American and Asian groups — on the Fair-Speech dataset expanded to more than twice its original size.
Assuming a correction cost of 5 seconds per transcription error, the correction time per minute of speech after pruning rose from 30 seconds to 64 seconds, a relative increase of +111%.
Even when using beam search decoding as mitigation, the group temporal tax gap caused by pruning still retained a +86% increase.
Wanda is a post-training pruning method that requires no retraining, determining which parameters can be zeroed out by combining weight magnitude with statistical information from input activations.
All Facts (4)
After applying 50% Wanda pruning to Whisper-large-v3, the absolute word error rate gap between the worst-served and best-served groups — specifically between Black/African American and Asian groups — on the Fair-Speech dataset expanded to more than twice its original size.
50%UnverifiedAssuming a correction cost of 5 seconds per transcription error, the correction time per minute of speech after pruning rose from 30 seconds to 64 seconds, a relative increase of +111%.
50%UnverifiedEven when using beam search decoding as mitigation, the group temporal tax gap caused by pruning still retained a +86% increase.
50%UnverifiedWanda is a post-training pruning method that requires no retraining, determining which parameters can be zeroed out by combining weight magnitude with statistical information from input activations.
50%