Summary of Wave: Weight Templates For Adaptive Initialization Of Variable-sized Models, by Fu Feng et al.
WAVE: Weight Templates for Adaptive Initialization of Variable-sized Models
by Fu Feng, Yucheng Xie, Jing Wang, Xin Geng
First submitted to arxiv on: 25 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: None
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The proposed WAVE approach tackles the issue of deploying pre-trained models in scenarios where target model sizes are incompatible with pre-trained ones. By reformulating variable-sized model initialization as a multi-task problem, WAVE employs shared weight templates and size-specific scalers to achieve consistent initialization across various model sizes. This is achieved through a distillation process constrained by Kronecker-based rules within the Learngene framework. The resulting initialized models demonstrate state-of-the-art performance in initializing models of varying depth and width. Furthermore, the knowledge encapsulated in weight templates is task-agnostic, enabling seamless transfer across diverse downstream datasets. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper introduces a new approach called WAVE that helps deploy pre-trained models in situations where the target model size doesn’t match the pre-trained one. It uses a multi-task method to initialize different-sized models and achieves this through a process that combines knowledge from pre-trained models with rules based on Kronecker products. The results show that this method can efficiently initialize models of varying sizes, and the learned knowledge can be applied to different tasks without needing much additional training. |
Keywords
» Artificial intelligence » Distillation » Multi task
