Expandable pruning: Enhancing model compression through smoothness-guided widening

Key Points

• Enhanced model compression is achieved through a novel smoothness-guided widening approach, allowing for better performance with fewer parameters.
• Key evidence shows that using smoothness-guided techniques results in a substantial reduction of model size while maintaining accuracy.
• The approach employs a unique method of expanding the width of neural networks, leveraging smoothness to enhance efficiency in parameter usage.
• This method highlights a potential shift in model training paradigms, suggesting that new strategies can optimize resource usage effectively.