KIF: Knowledge Identification and Fusion for Language Model Continual Learning

TL;DR

This paper introduces a novel CL framework for language models, named Knowledge Identification and Fusion (KIF), which boosts knowledge transfer without depending on memory replay, and demonstrates strong generalizability, making it suitable for various base models and adaptable to PEFT methods like LoRA.

Abstract

Language model continual learning (CL) has recently attracted significant interest for its ability to adapt large language models (LLMs) to dynamic real-world scenarios without retraining. A major challenge in this domain is catastrophic forgetting, where models lose previously acquired knowledge upon learning new tasks. Existing approaches commonly utilize multiple parameter-efficient fine-tuning (PEFT) blocks to acquire task-specific knowledge, yet these methods are inefficient and fail to leverage potential knowledge transfer across tasks. In this paper, we introduce a novel CL framework for language models, named Knowledge Identification and Fusion (KIF), which boosts knowledge transfer without depending on memory replay. KIF initially segregates the model into 'skill units' based on parameter dependencies, allowing for more precise control. Subsequently, it employs a novel group-wise knowledge identification technique to ascertain the importance distribution of skill units for a new task. By comparing this importance distribution with those from previous tasks, we implement a fine-grained knowledge fusion strategy that retains task-specific knowledge, thereby preventing forgetting, and updates task-shared knowledge, which facilitates bi-directional knowledge transfer. As a result, KIF achieves an optimal balance between retaining prior knowledge and excelling in new tasks. KIF also demonstrates strong generalizability, making it suitable for various base models and adaptable to PEFT methods like LoRA. Furthermore, it offers notable extensibility, supporting enhancements through integration with memory replay techniques. Comprehensive experiments conducted on two CL benchmarks, involving models ranging from 220M to 7B parameters, affirm the effectiveness of KIF and its variants across different settings.

Figures (7)

• Figure 1: Conceptual illustration of KIF. By identifying task-relevant areas across both previously accumulated and current tasks, we can consolidate the task-shared and task-specific parameters to facilitate efficient knowledge transfer (KT) and mitigate catastrophic forgetting (CF).
• Figure 2: Left: Depiction of reconstructing LoRA as fine-grained skill units. Right:Overview of KIF.Step 1: We compute the importance scores of skill units for the current task $k$ using our importance-aware knowledge identification method during fine-tuning. Step 2: Based on a categorical model averaging strategy, the knowledge fusion mechanism merges the model $\hat{f}_{k-1}$, which accumulates knowledge of all previous tasks, with the current task's model $f_k$. This integration is strategically guided by the importance distributions of skill units across different tasks. This iterative process continues with the addition of each new task.
• Figure 3: Overview of KIF-M. By fixing the current task-specific skill unit during the replay phase, we can further enhance the performance of KIF.
• Figure 4: Performance of KIFLoRA across various backbones.
• Figure 5: Performance trajectory of Task 1 on Long Sequence Benchmark throughout the continual learning process.