Saved in:
Bibliographic Details
Main Authors: Diecidue, Andrea, Barbano, Carlo Alberto, Fraternali, Piero, Fontaine, Mathieu, Tartaglione, Enzo
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2509.26207
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866918390987227136
author Diecidue, Andrea
Barbano, Carlo Alberto
Fraternali, Piero
Fontaine, Mathieu
Tartaglione, Enzo
author_facet Diecidue, Andrea
Barbano, Carlo Alberto
Fraternali, Piero
Fontaine, Mathieu
Tartaglione, Enzo
contents Transformer-based models have become the state of the art across multiple domains, from natural language processing to machine listening, thanks to the attention mechanisms. However, the attention layers require a large number of parameters and high-end hardware for both training and inference. We propose a novel channel-pruning technique explicitly targeted at the attention mechanism, decoupling the pruning of each head and the four layers in the attention block: query, key, value, and output projection matrices, employing a second-order metric to score the network's parameters. We compare our technique against head-pruning strategies and magnitude-driven scoring metrics, investigating the effects of pruning on Audio Spectrogram Transformer (AST) and Whisper. Our results show that even after pruning 50\% of the parameters in the attention block, performance is largely preserved.
format Preprint
id arxiv_https___arxiv_org_abs_2509_26207
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle The silence of the weights: a structural pruning strategy for attention-based audio signal architectures with second order metrics
Diecidue, Andrea
Barbano, Carlo Alberto
Fraternali, Piero
Fontaine, Mathieu
Tartaglione, Enzo
Sound
Machine Learning
Transformer-based models have become the state of the art across multiple domains, from natural language processing to machine listening, thanks to the attention mechanisms. However, the attention layers require a large number of parameters and high-end hardware for both training and inference. We propose a novel channel-pruning technique explicitly targeted at the attention mechanism, decoupling the pruning of each head and the four layers in the attention block: query, key, value, and output projection matrices, employing a second-order metric to score the network's parameters. We compare our technique against head-pruning strategies and magnitude-driven scoring metrics, investigating the effects of pruning on Audio Spectrogram Transformer (AST) and Whisper. Our results show that even after pruning 50\% of the parameters in the attention block, performance is largely preserved.
title The silence of the weights: a structural pruning strategy for attention-based audio signal architectures with second order metrics
topic Sound
Machine Learning
url https://arxiv.org/abs/2509.26207