Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

FGGM: Fisher-Guided Gradient Masking for Continual Learning

Chao-Hong Tan, Qian Chen, Wen Wang, Yukun Ma, Chong Zhang, Chong Deng, Qinglin Zhang, Xiangang Li, Jieping Ye

TL;DR

FGGM addresses catastrophic forgetting in continual learning for large language models by introducing Fisher-guided gradient masking, which uses the diagonal Fisher Information Matrix to identify and protect critical parameters while updating others. The method operates without access to historical data, generating binary masks with adaptive thresholds and applying gradient projection to constrain updates, thereby balancing stability and plasticity. Empirical results on TRACE show FGGM outperforming prior methods like MIGU and achieving strong performance on code-generation benchmarks, with clear ablations highlighting the importance of input-dimension aggregation. The approach offers a principled, memory-efficient solution for no-data continual learning in LLMs, with scalability to larger models and tasks.

Abstract

Catastrophic forgetting impairs the continuous learning of large language models. We propose Fisher-Guided Gradient Masking (FGGM), a framework that mitigates this by strategically selecting parameters for updates using diagonal Fisher Information. FGGM dynamically generates binary masks with adaptive thresholds, preserving critical parameters to balance stability and plasticity without requiring historical data. Unlike magnitude-based methods such as MIGU, our approach offers a mathematically principled parameter importance estimation. On the TRACE benchmark, FGGM shows a 9.6% relative improvement in retaining general capabilities over supervised fine-tuning (SFT) and a 4.4% improvement over MIGU on TRACE tasks. Additional analysis on code generation tasks confirms FGGM's superior performance and reduced forgetting, establishing it as an effective solution.

FGGM: Fisher-Guided Gradient Masking for Continual Learning

TL;DR

FGGM addresses catastrophic forgetting in continual learning for large language models by introducing Fisher-guided gradient masking, which uses the diagonal Fisher Information Matrix to identify and protect critical parameters while updating others. The method operates without access to historical data, generating binary masks with adaptive thresholds and applying gradient projection to constrain updates, thereby balancing stability and plasticity. Empirical results on TRACE show FGGM outperforming prior methods like MIGU and achieving strong performance on code-generation benchmarks, with clear ablations highlighting the importance of input-dimension aggregation. The approach offers a principled, memory-efficient solution for no-data continual learning in LLMs, with scalability to larger models and tasks.

Abstract

Catastrophic forgetting impairs the continuous learning of large language models. We propose Fisher-Guided Gradient Masking (FGGM), a framework that mitigates this by strategically selecting parameters for updates using diagonal Fisher Information. FGGM dynamically generates binary masks with adaptive thresholds, preserving critical parameters to balance stability and plasticity without requiring historical data. Unlike magnitude-based methods such as MIGU, our approach offers a mathematically principled parameter importance estimation. On the TRACE benchmark, FGGM shows a 9.6% relative improvement in retaining general capabilities over supervised fine-tuning (SFT) and a 4.4% improvement over MIGU on TRACE tasks. Additional analysis on code generation tasks confirms FGGM's superior performance and reduced forgetting, establishing it as an effective solution.
Paper Structure (11 sections, 6 equations, 2 figures, 3 tables)

This paper contains 11 sections, 6 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. The FGGM Framework
  4. Overview
  5. Fisher-Guided Mask Initialization
  6. Mask Normalization via Output Dimension
  7. Gradient Projection
  8. Experiments
  9. Experimental Setup
  10. Results and Analysis
  11. Conclusion

Figures (2)

  • Figure 1: Illustration of the FGGM training pipeline.
  • Figure 2: Performance comparison between SFT, MIGU, and our FGGM on the HumanEval and BBH benchmarks through progression of training epoches, from training on Magicoder-Eval-Instruct-110K, all in 0.5B model size.