Tied-Lora: Enhancing parameter efficiency of LoRA with weight tying

TL;DR

Tied-LoRA introduces weight tying and selective training to improve parameter efficiency in Low-rank Adaptation, achieving comparable performance with fewer trainable parameters across multiple tasks and models.

Contribution

The paper proposes Tied-LoRA, a novel method combining weight tying with selective training to enhance parameter efficiency in LoRA.

Findings

Tied-LoRA achieves similar performance to LoRA with fewer parameters.

A specific Tied-LoRA configuration outperforms standard LoRA at high ranks.

Experiments across diverse tasks validate the efficiency-performance trade-off.

Abstract

We introduce Tied-LoRA, a novel paradigm leveraging weight tying and selective training to enhance the parameter efficiency of Low-rank Adaptation (LoRA). Our exploration encompasses different plausible combinations of parameter training and freezing, coupled with weight tying, aimed at identifying the optimal trade-off between performance and the count of trainable parameters. Across $5$ diverse tasks and two foundational language models with different parameter counts, our experiments provide comprehensive insights into the inherent trade-offs between efficiency and performance. Our findings reveal a specific Tied-LoRA configuration that distinguishes itself by showcasing comparable performance to LoRA across multiple tasks while utilizing only a fraction of the parameters employed by the standard LoRA method, particularly at elevated ranks. This underscores the efficacy of Tied-LoRA in achieving impressive results with significantly reduced model complexity.