Reproduce “GRAM: Fast Fine-tuning of Pre-trained Language Models for Content-based Collaborative Filtering” on OpenResearch

About this paper

Yoonseok Yang, Kyu Seok Kim, Minsam Kim, Juneyoung Park

Riiid AI Research developed GRAM, a training framework that optimizes fine-tuning of pre-trained language models for content-based collaborative filtering by intelligently accumulating gradients for unique items. This approach significantly reduces training time and GPU memory, demonstrating up to 146x speedup and enabling the use of richer content features with minimal performance degradation compared to end-to-end training.

Key insights

• The core inefficiency in fine-tuning pre-trained language models for content-based collaborative filtering arises from redundant gradient computations for the same item content across a batch.
• Aggregating gradients for only the unique items within a batch or over multiple steps can significantly reduce computational load without substantial performance degradation.
• GRAM's approach is orthogonal to existing efficiency methods like parameter-efficient transfer learning or knowledge distillation, allowing for combined benefits.

Summary by alphaXiv.