Space potential measurement method and system for neural network small sample learning

Abstract

The invention provides a space potential measurement method for small sample learning of a neural network, which comprises the following steps of: firstly, respectively carrying out feature extraction on a support set and a query set in a certain scene through a feature embedding extraction network f theta to obtain feature embedding vector sets of marked instances and unmarked instances; secondly, performing parameter extraction on the feature embedding vectors of the marked instances through a charge quantity parameter extraction network g theta to obtain charge quantity range parameters of the marked instances, and performing similar mean value fusion; then, the feature embedding vector set of the marked instances and the charge quantity range parameters of the marked instances are utilized to construct a quasi-space electrostatic field, and the superimposed potential of the unmarked instances in the electrostatic field is used as a metric value between the unmarked instances and the marked instances; and finally, obtaining a category probability and learning loss according to the metric value, and updating all network parameters in a back propagation manner. The method provided by the invention can be compatible with the vast majority of small sample networks at lower model cost, and further improves the expressive force of small sample tasks.

Description

technical field [0001] The invention relates to the field of machine learning, in particular to a space potential measurement method and system for neural network small-sample learning. Background technique [0002] In recent years, deep learning models have made major breakthroughs in computer vision tasks, such as image classification, semantic segmentation, and object detection. By training and iterating a large number of data samples, the model can perform even high to human capabilities. In most learning tasks, the prerequisite for deep learning to show superior performance is often to have a large number of labeled samples. When the learning task changes slightly, the results usually collapse. In contrast, humans can generalize important knowledge from a small number of examples and apply it to new scenarios, which we refer to as the ability to learn to learn, or meta-learning. [0003] Small-sample learning is a major application in the field of meta-learning, which...

AI Technical Summary

Problems solved by technology

[0008] The main purpose of the present invention is to overcome the above-mentioned defects in the prior art, and propose a space potential measurement method for neural network small-sample learning, thereby solving the problem that the existing small-sample measurement technology cannot utilize low model cost, low Computational complexity comprehensively considers the technical issues of the relationship within categories and between categories

Images