Sparse Adapter Fusion for Continual Learning in NLP

TL;DR

SAFM addresses catastrophic forgetting in NLP by sparsely fusing adapters across tasks and selectively reusing existing modules or introducing empty adapters to minimize parameter growth. It introduces a two-stage workflow—decision stage with architecture search and tuning stage with pseudo-replay and a layer-wise loss—to balance global knowledge sharing and local task specificity. The approach yields competitive or superior performance while using less than 60% of the parameters of SOTA methods and demonstrates positive backward transfer across diverse task scenarios and backbones. This work offers a practical, scalable solution for parameter-efficient continual learning in NLP, with clear benefits for cross-task knowledge transfer and model efficiency.

Abstract

Continual learning in natural language processing plays a crucial role in adapting to evolving data and preventing catastrophic forgetting. Despite significant progress, existing methods still face challenges, such as inefficient parameter reuse across tasks, risking catastrophic forgetting when tasks are dissimilar, and the unnecessary introduction of new parameters for each task, which hampers knowledge sharing among similar tasks. To tackle these issues, we propose a Sparse Adapter Fusion Method (SAFM), which dynamically fuses old and new adapters to address these challenges. SAFM operates in two stages: the decision stage and the tuning stage. In the decision stage, SAFM determines whether to incorporate a new adapter, reuse an existing one, or add an empty adapter. The architecture search procedure, designed to prioritize reusing or adding empty adapters, minimizes parameter consumption and maximizes reuse. In the tuning stage, SAFM especially facilitates a layer-wise loss to encourage differentiation between adapters, effectively capturing knowledge within the same task. Experimental results consistently show that SAFM outperforms state-of-the-art (SOTA) methods, achieving comparable performance while utilizing less than 60% of the parameters.

Figures (2)

• Figure 1: SAFM consists of two stages: the decision stage and the tuning stage. The color indicates a module is updated with data from a specific task: green for task ${\mathcal{T}}_1$, yellow for task ${\mathcal{T}}_{n-2}$, and blue for task ${\mathcal{T}}_n$. In the decision stage, a new module (blue) is initialized for task ${\mathcal{T}}_n$, and the corresponding architecture search procedure is determined by Eq. (\ref{['eq:lambda']}), which yields $A_n^l=A_1^l=A_{n-2}^l$. Hence, at the tuning stage, $A_n^l$ has to be fine-tuned with data from task ${\mathcal{T}}_1$ (green), ${\mathcal{T}}_{n-2}$ (yellow), and ${\mathcal{T}}_n$ (blue). We then generate pseudo-samples from ${\mathcal{T}}_1$ ($P_1$) and ${\mathcal{T}}_{n-2}$ ($P_{n-2}$) with the incoming data in ${\mathcal{D}}_n$ to update the module $A_n^l$. For further details, please refer to Sec. \ref{['sec:archi']}.
• Figure 2: Learning curve of compared methods. SAFM's position above the other lines indicates its superior performance.