论文部分内容阅读
机器学习模型的生命周期通常比较短暂,大量的机器学习模型针对特定任务设计,在完成任务之后即失去使用价值.然而,一个精心设计和训练的模型通常更精炼地概括了训练数据中蕴含的知识.更进一步地,当无法获取原始训练数据时,已有的预训练模型就是仅剩的信息来源.本文提出了一种重用已有的预训练机器学习模型来辅助当前任务学习的框架,尤其适用于当前任务数据量不足的情形.该方法把已有模型视为黑盒,不需要已有模型的结构信息,并使用领域知识对已有模型进行筛选和结合.实验表明,在该框架下,通过对已有模型的重用,多个实际任务上的性能可以得到显著提升.
The life cycle of a machine learning model is often short lived and a large number of machine learning models are task-specific designed to lose use value after a task is completed.However, a well-designed and trained model generally provides a more refined overview of the knowledge contained in the training data Furthermore, when the original training data can not be obtained, the existing pre-training model is the only source of information.This paper presents a framework that reuses the existing pre-training machine learning model to assist the current task learning, especially for In the case of insufficient data amount of the current task, this method regards the existing model as a black box, does not need the structural information of the existing model, and uses the domain knowledge to filter and combine the existing models.Experiments show that under this framework, Through the reuse of existing models, the performance of multiple real tasks can be significantly improved.