Newly-added category detection method based on multi-sample learning

Abstract

The invention discloses a newly-added category detection method based on multi-sample learning. Firstly, key samples in a multi-sample 'package' are extracted through a relatively mature key sample detection algorithm in multi-sample learning; secondly, the key samples corresponding to each known category are combined into a 'category super package', and all samples not identified as key samples constitute a 'meta super package'; and thirdly, the distance between a package and a super package can be determined through subsequent metric learning. In the practical stage, for a package of a known category, the concept category of the package is determined according to a category super package nearest to the package; and for a package of a newly-added category, as there is no category super package corresponding to the package and a super package nearest to the package is a meta super package, the category of the package can be determined as a newly-added category.

Description

technical field [0001] The present invention relates to machine learning and application technology, in particular to multi-instance learning, new category monitoring technology, and metric learning. It is an automatic concept category prediction / classification of existing categories and detection of new categories Robust multiple-instance learning algorithm. Background technique [0002] Learning from examples is considered to be the most promising approach to machine learning. If the ambiguity of the training samples is used as the division criterion, the current research in this field is roughly based on three learning frameworks, namely supervised learning, unsupervised learning and reinforcement learning. [0003] Supervised learning works by learning on training examples with concept labels in order to predict concept labels for examples outside the training set as correctly as possible. All training samples here are labeled, so their ambiguity is the lowest. Unsupe...

AI Technical Summary

Problems solved by technology

In the practical stage of the system, it is very likely that new image categories will appear, such as "tiger"

At this time, the existing multi-instance learning algorithm can only simply misclassify the samples belonging to the newly added category (such as pictures of tigers) into a certain category of known categories (such as "fox"), which will make the system Fails in a dynamic open environment

Images