Reproduce “LangVision-LoRA-NAS: Neural Architecture Search for Variable LoRA Rank in Vision Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Krishna Teja Chitty-Venkata, Murali Emani, Venkatram Vishwanath
Researchers at Argonne National Laboratory developed LangVision-LoRA-NAS, a framework that integrates Neural Architecture Search (NAS) with Low-Rank Adaptation (LoRA) to automatically optimize layer-wise LoRA ranks for Vision-Language Models (VLMs). This approach achieved a 2.6x reduction in trainable LoRA parameters and reduced fine-tuning epoch times while maintaining or slightly improving model performance across ten diverse multimodal tasks.
Key insights
- Integrating NAS with LoRA enables dynamic, layer-wise optimization of LoRA ranks in VLMs, moving beyond fixed-rank paradigms to balance performance and efficiency.
- Using a Supernetwork with differentiable architectural parameters allows for efficient, gradient-based search for optimal LoRA ranks without needing to train multiple models.
- Optimal LoRA ranks are not uniform across VLM layers and modules, suggesting that tailoring adaptation capacity to specific components yields better resource utilization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.