Redirigiendo al acceso original de articulo en 24 segundos...
Inicio  /  Algorithms  /  Vol: 16 Par: 9 (2023)  /  Artículo
ARTÍCULO
TITULO

Variable Scale Pruning for Transformer Model Compression in End-to-End Speech Recognition

Leila Ben Letaifa and Jean-Luc Rouas    

Resumen

Transformer models are being increasingly used in end-to-end speech recognition systems for their performance. However, their substantial size poses challenges for deploying them in real-world applications. These models heavily rely on attention and feedforward layers, with the latter containing a vast number of parameters that significantly contribute to the model?s memory footprint. Consequently, it becomes pertinent to consider pruning these layers to reduce the model?s size. In this article, our primary focus is on the feedforward layers. We conduct a comprehensive analysis of their parameter count and distribution. Specifically, we examine the weight distribution within each layer and observe how the weight values progress across the transformer model?s blocks. Our findings demonstrate a correlation between the depth of the feedforward layers and the magnitude of their weights. Consequently, layers with higher weight values require less pruning. Building upon this insight, we propose a novel pruning algorithm based on variable rates. This approach sets the pruning rate according to the significance and location of each feedforward layer within the network. To evaluate our new pruning method, we conduct experiments on various datasets. The results reveal its superiority over conventional pruning techniques, such as local pruning and global pruning.

 Artículos similares

       
 
Yaneth Vasquez, Jair Franco, Mario Vasquez, Felipe Agudelo, Eleni Petala, Jan Filip, Jose Galvis and Oscar Herrera    
The tannery wastewater from the tanning stage (TWT) comprises organic and Cr pollutants, which can adversely affect aquatic life and have carcinogenic effects. In this study, we investigated the performance of a Fenton-like process using commercial Nano-... ver más
Revista: Water

 
Huimin Li and Yijun He    
Spaceborne synthetic aperture radar (SAR) has been widely acknowledged for its advantages in collecting ocean surface measurements under all weather conditions during day and night. Despite the strongly nonlinear imaging process, SAR measurements of ocea... ver más

 
Jiaqi Hu, Yin Gu, Jinhuang Yan, Ying Sun and Xinyi Huang    
With the convenient and fast requirements for construction in bridge engineering, prefabricated assembly technology is widely applied in engineering construction. Typically, prefabricated bridge decks are connected through cast-in-place wet joints. Wet j... ver más
Revista: Applied Sciences

 
Jennifer Coston-Guarini, François Charles and Jean-Marc Guarini    
An outbreak species exhibits extreme, rapid population fluctuations that can be qualified as discrete events within a continuous dynamic. When outbreaks occur they may appear novel and disconcerting because the limiting factors of their dynamics are not ... ver más

 
Mahsa Kashani, Stefan Jespersen and Zhenyu Yang    
The application of deoiling hydrocyclone systems as the downstream of three-phase gravity separator (TPGS) systems is one of the most commonly deployed produced water treatment processes in offshore oil and gas production. Due to the compact system?s com... ver más
Revista: Water