Multi-task and cross-task supporting small sample classification training method and device

Abstract

The invention discloses a multi-task and cross-task supporting small sample classification training method and device, and the method comprises the steps: carrying out the formalized analogy: enablinga classification task in a small sample classification problem to be converted into a sample in a standard classification problem, learning a task solver (capable of estimating whether a task is completed or not) under the condition of forming a target form of small sample classification into a given large number of task samples; 2) simulating a batch training technology in a standard classification problem (processing some samples in each category in each iteration), and proposing a small sample classification training algorithm of multitask (processing some task samples in multiple task categories in each iteration); and 3) simulating a pre-training technology in a standard classification problem (a basic model is pre-trained for a similar small-scale data task on large-scale data), andproposing a cross-task small sample classification training algorithm (a basic model is pre-trained for a small-class (low-class) problem on a multi-class (high-class) problem).

Description

technical field

[0001] The embodiments of the present application relate to deep learning, image classification and computer vision processing technologies, and in particular to a small-sample classification training method and device supporting multi-task and cross-task. Background technique

[0002] In recent years, thanks to the development of deep learning technology, breakthroughs have been made in large-scale supervised learning, especially in the field of image recognition. For example, the accuracy on the ImageNet dataset has increased from 50% in 2012 to 80%. The accuracy of face recognition even exceeds that of human eyes. But behind the success of deep learning is the dependence on large data sets. In reality, for example, in the automatic identification of traffic accidents, the classification of military sensitive targets, and the toxicity testing of pharmaceutical molecules, the samples that can be obtained are very scarce. At this time, directly using tradit...

Images