Self-supervised learning method and self-supervised learning device

Abstract

The invention provides a self-supervised learning method and a self-supervised learning device, and relates to the field of computers. According to the invention, through the distance measurement information from the feature of each augmented picture of each original picture to the feature mean value of all augmented pictures of each original picture, the augmented picture sample with semantic deviation can be effectively monitored, and the corresponding weight is reduced, so that the augmented picture sample can be effectively monitored. The training noise brought by the augmented picture sample with semantic deviation can be effectively suppressed, the variance and deviation of data augmented distribution are well balanced, and the performance of a learning model on a downstream computer vision task is improved.

Description

technical field [0001] The disclosure relates to the field of computers, in particular to a self-supervised learning method and a self-supervised learning device. Background technique [0002] Self-supervised learning is a very effective feature learning method. Self-supervised learning assumes that the category labels of the training data are unknown, and under this premise, the characteristics of the training data are learned through the structural assumptions of the training data itself. After unsupervised pre-training of deep networks using self-supervised learning methods, deep networks can often achieve very good results in downstream computer vision tasks such as classification, monitoring, and segmentation. [0003] There is a self-supervised learning method that artificially creates various augmented images of the original image by flipping, translating, adding noise, cutting, etc., and assumes that these augmented images of the same original image have the consist...

AI Technical Summary

Problems solved by technology

[0004] However, these randomly generated augmented images may have deviated from the assumption of semantic invariance at the feature space level, and do not have the characteristics of semantic invariance

For example, over-intensive data augmentation of an original image may cause the resulting augmented image to be semantically biased

If such biases are ignored, forcing deep networks to learn semantically invariant semantics implicit in augmented images will negatively impact the performance of downstream computer vision tasks.

Images