This paper presents a comprehensive survey of the latest advancements in pre-trained model-based continual learning. The authors categorize existing methodologies into three distinct groups: prompt-based methods, representation-based methods, and model mixture-based methods.
Prompt-based methods utilize lightweight prompts to adapt pre-trained models to new tasks while preserving the generalizability of the pre-trained backbone. These methods focus on prompt selection and combination techniques to enable instance-specific knowledge retrieval.
Representation-based methods directly leverage the strong representational capabilities of pre-trained models, constructing classifiers based on the frozen backbone features. These methods demonstrate the inherent ability of pre-trained models to generalize to new tasks.
Model mixture-based methods create a set of models during the continual learning process and employ model ensemble or model merge techniques to derive the final prediction. This approach aims to alleviate catastrophic forgetting by preserving knowledge from multiple learning stages.
The authors provide an empirical study contrasting various state-of-the-art methods across seven benchmark datasets, highlighting concerns regarding fairness in comparisons. They also discuss future directions, including the potential of continual learning with large language models, expanding beyond single-modality recognition, and the need for new benchmarks that challenge the knowledge of pre-trained models.
To Another Language
from source content
arxiv.org
Дополнительные вопросы