Image classification method and device based on meta learning, product and storage medium

Abstract

The invention discloses an image classification method and device based on meta learning, a product and a storage medium. The method comprises the following steps: training a sample feature extractor and a task feature extractor F for a current task according to an input training sample, and obtaining a corresponding task feature vi; inputting the vi into a clustering network layer, updating a corresponding clustering feature vector set through minimum distance matching and an online clustering mode, and outputting a clustering feature vector c[omega] which is most adaptive to the current task; performing correlation evaluation on the task feature vector c[omega] and the current task feature vi based on the task feature evaluation network sigma, outputting a corresponding evaluation coefficient [alpha], and calculating a representation vector [omega]out suitable for the current task based on the coefficient; modulating the parameters of the model based on the [omega]out to obtain an optimal model initialization parameter [theta]new; and enabling the model to start to train and update based on the initialization parameter [theta]new. Updating of task priori knowledge, correlation adaptation and efficient multiplexing can be completed under the condition that man-made interference is not carried out, and learning of a new image classification task is accelerated.

Description

technical field [0001] The invention relates to the fields of meta-learning and image recognition, in particular to an image classification method, device, product and storage medium based on meta-learning. Background technique [0002] In recent years, the development of the field of deep learning has brought machine learning to a new stage, and the development of deep neural networks has also enabled machine learning models to achieve excellent performance. However, most of the existing machine learning models require a large number of labeled training samples and a large amount of training time. However, corresponding to real life, the label collection of samples is difficult, and sometimes there are differences in the richness of the samples themselves, and some tasks may only have a small number of training samples. At the same time, for time-sensitive model training, it is not allowed to use a lot of time for training. In view of the above problems, one solution is t...

AI Technical Summary

Problems solved by technology

There are two main defects in the above method: ①The model does not consider whether the new task is related to the task learned in the past, and the initialization parameters are only related to the task learned in the past

② The tasks learned by the model in the past may have differences in their distribution. Generating the best model initialization parameters corresponding to all tasks may make the model lack specificity, resulting in poor performance on new tasks

The method of multiple classifiers enables new evaluation data to use more effective classifiers, but it also limits the scalability of the model. If there is a large difference between the new prediction task and the task used in preprocessing, then Classifiers based on pre-training may not be suitable for new tasks

Images