Zero sample learning method based on global semantic consistency network

Abstract

The invention belongs to the technical field of machine learning, and in particular relates to a zero sample learning method based on a global semantic consistent network. The semantic information ofthe whole class act as the weight of the full connection layer to be added to the deep learning framework, and the nonlinear self-weighting structure of the corresponding product of two full connection corresponding products and the cross entropy loss function are applied to establish the global semantic consistent network GSC-Net and end-to-end zero-sample learning is implemented through GSC-Net.The method is simple in framework, convenient in use, high in expansibility and high in interpretability and far exceeds the existing method in the result of two major tasks of zero sample classification and generalized zero sample classification of three mainstream visual attribute data sets. The method can provide the basic framework and algorithm support for computer vision, natural language processing, recommendation system and other fields related to zero sample learning and can be easily extended to problems such as open set identification, incremental learning, online learning and thelike.

Description

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine

Login to view more

AI Technical Summary

Benefits of technology

This patented technology allows us to accurately identify unknown objects based on their appearance features or properties. It uses convolution networks that process images from multiple sources together into an ensemble representation called vector space representations (VSM). These VSF models represent how similar things appear differently depending upon factors like color intensity, shape, texture etc., making them ideal tools for object categorization purposes such as machine learning applications. Overall this technology provides technical benefits over current techniques including binary cross-correlation and histogram analysis.

Problems solved by technology

Technological Problem addressed in this patents relates to improving the performance of machine learning algorithms used during real world scenarios like biomedical imagery analysis. Current solutions either require large amounts of annotated dataset or lack robust annotation labels without addressing issues related to imprecision sampling or non-uniformly assigning representative examples across multiple domains.

Images