Core Concepts
Q-tuning, a novel approach for continual prompt tuning, enables lifelong learning of a pre-trained language model by managing a prompt queue and adaptively aggregating previous prompts.
Abstract
The paper introduces Q-tuning, a novel approach for continual prompt tuning that enables lifelong learning of a pre-trained language model.
Key highlights:
Q-tuning manages a prompt queue (Q-prompt) that stores previously learned prompts. For a new task, Q-tuning trains a new prompt combined with the fixed Q-prompt.
Q-tuning uses an adaptive knowledge aggregation technique to reweigh previous prompts in the queue, enhancing forward knowledge transfer.
When the Q-prompt reaches its maximum capacity, Q-tuning leverages a PCA-based eviction rule to reduce the queue size while preserving primary knowledge of old tasks.
To mitigate information loss from eviction, Q-tuning proposes a globally shared prefix prompt and a memory retention regularization.
Extensive experiments demonstrate Q-tuning outperforms state-of-the-art continual learning and prompt tuning methods, especially on long task sequences that mimic lifelong learning scenarios.
Stats
The paper reports accuracy metrics on various few-shot continual learning benchmarks with short and long task sequences.
Quotes
"Q-tuning manages a Queue-based prompt (Q-prompt), which is stored in a finite-size data buffer."
"We design an adaptive knowledge aggregation technique that reweighs previous prompts in the queue with a learnable low-rank matrix."
"We leverage a PCA-based eviction rule to reduce the queue's size, allowing the newly trained prompt to be added while preserving the primary knowledge of old tasks."