Efficient Fine-Tuning of Large Language Models

The field of large language models is moving towards more efficient fine-tuning methods, with a focus on reducing computational costs and parameter sizes. Recent developments have introduced novel approaches to parameter-efficient transfer learning, such as integrating shared and layer-specific information, utilizing low-rank symmetric weight matrices, and leveraging Fisher information to select critical parameters. These innovations have led to significant improvements in performance while maintaining superior parameter efficiency. Noteworthy papers include:

  • Optimizing Specific and Shared Parameters for Efficient Parameter Tuning, which proposes a novel PETL method that effectively mitigates distributional shifts during fine-tuning.
  • FISH-Tuning, which incorporates FISH Mask into addition-based and reparameterization-based PEFT methods to achieve superior performance without additional memory overhead or inference latency.
  • AROMA, which introduces a dual-loop architecture for rank growth and significantly reduces parameters compared to LoRA and AdaLoRA while achieving superior performance on natural language understanding and commonsense reasoning tasks.

Sources

Optimizing Specific and Shared Parameters for Efficient Parameter Tuning

Towards Symmetric Low-Rank Adapters

FISH-Tuning: Enhancing PEFT Methods with Fisher Information

AROMA: Autonomous Rank-one Matrix Adaptation

Less but Better: Parameter-Efficient Fine-Tuning of Large Language Models for Personality Detection

Built with on top of