Variational auto-encoder network model and variational auto-encoder network model device for small sample learning

Abstract

The invention provides a variational auto-encoder network model and device for small sample learning, and the method comprises the steps: obtaining the high-level representation of a sample, obtainingthe implicit representation through the sampling of a first neural network according to the high-level representation, and restoring the implicit representation through a second neural network, so asto reconstruct the high-level representation; generating each category distribution according to the distribution parameters; and defining a loss function as the sum of the reconstruction error, theclassification error, the implicit representation distribution and the priori distribution error to optimize the whole network parameters and the parameters of each category distribution, and generating a variational auto-encoder network model. Through the variational auto-encoder network model oriented to small sample learning, the learned high-level representation can be accurately described bya distribution family, and the distribution of the high-level representation can be completely matched with the metric function of the high-level representation, so that the accuracy of small sample classification is improved.

Description

technical field [0001] The invention relates to the technical field of small-sample machine learning, in particular to a variational autoencoder network model and device for small-sample learning. Background technique [0002] At present, small sample machine learning solves how to train a more stable and better classifier with a limited number of training samples. Generally speaking, in small-sample deep learning, there is a set of base classes with sufficient samples to learn "Meta Knowledge", and then generalize it to small-sample new classes. [0003] In the prior art, the method of combining metric learning (Metric Learning) with deep network end-to-end training has achieved good results. However, the high-level features of samples learned by deep learning generally do not have a distribution family that can accurately describe them, and a major shortcoming is that the measurement function does not completely match the distribution of high-level features of samples, th...

AI Technical Summary

Problems solved by technology

However, the high-level features of samples learned by deep learning generally do not have a distribution family that can accurately describe them, and a major shortcoming is that the measurement function does not completely match the distribution of high-level features of samples, thus causing a certain degree of loss of precision

Images