Training sample annotation cost reduction method for transfer learning

Abstract

The invention discloses a training sample annotation cost reduction method for transfer learning. The method is used for carrying out distinguishing processing on a to-be-annotated sample of a targettask. The method comprises the following steps: annotating a target task sample which is not located in a source task sample space; and for a target task sample located in the source task sample space, automatically finding the optimal substitution from the labeled sample set of the source task. According to the method, the annotation quantity of the training samples can be reduced, repeated annotation of the samples is avoided, the annotation sample quantity is increased under the condition that the annotation cost is not increased, and the actual sample annotation cost can be controlled by adjusting parameters. Besides, the method can keep the model performance stable under the condition of compressing a large annotation cost, and can be matched with any sample selection method for use.Meanwhile, the method is also suitable for application scenarios in which the model needs to be updated and a personalized model is established on the basis of a universal model.

Description

technical field [0001] The invention relates to the field of migration learning of artificial intelligence, in particular to a method for reducing the labeling cost of target task samples by utilizing source task knowledge and existing labeling data. Background technique [0002] Current successful machine learning models attach great importance to data, and their high performance relies on a large amount of labeled data. Machine learning models have been successfully applied to tasks that have accumulated or can easily obtain a large amount of labeled data (such as computer vision, text translation, speech recognition, etc.), however, more application scenarios have not accumulated or can not easily obtain a large amount of labeled data. To cope with the long tail of the distribution of application scenarios, the task knowledge that has been learned needs to be transferable to new tasks. [0003] The goal of transfer learning is to transfer the knowledge learned from the s...

AI Technical Summary

Problems solved by technology

In this way, a large number of labeled samples can be obtained at one time, but there are the following problems: (1) The information of these source task samples has already been included in some models and parameters from the source task used by the target model, and it is somewhat difficult to use them again for training. waste of computing resources

(2) It does not solve the problem of repeated labeling waste when labeling new target task samples

(4) Due to the difference between the source task and the target task, a large number of different samples ( figure 1 samples in region 1) will affect the learning of the real target task

(2) The sample in area 2 is retrieved as a substitute for the current target task sample, but the retrieved sample cannot truly represent the current target task sample

[0010] The choice of public feature space will obviously affect the effect of final model learning, but finding the best public feature space is transfer learning rather than the concern of the present invention

Images