Multi-task learning method and system based on feature and sample adversarial symbiosis

Abstract

The invention relates to the field of multi-task deep learning, and provides a multi-task learning method based on feature and sample adversarial symbiosis, which comprises the following steps: S1, randomly extracting samples of tasks, and generating common implicit features irrelevant to the field; s2, based on the common implicit characteristics generated in the step S1, generating a high-simulation sample, and taking the high-simulation sample as a task sample of the next cycle in the step S1; s3, circulating the steps S1 and S2 until the multi-task adversarial game is balanced, and generating a final high-simulation sample and a high-quality classification label. The invention further provides a multi-task learning system based on feature and sample adversarial symbiosis. According tothe method, the problems of domain distribution difference and small samples are solved, and the generalization performance of a machine learning system is greatly improved, so that a plurality of application fields of artificial intelligence are promoted to be broken through. The method is not only suitable for multi-task learning and transfer learning, but also suitable for multi-view learning and multi-modal learning.

Description

technical field [0001] The invention relates to the field of multi-task deep learning, in particular to a multi-task learning method and system based on feature and sample confrontation symbiosis. Background technique [0002] Deep learning originated from the research of artificial neural network, and it has outstanding performance in the application of many fields of artificial intelligence. Deep learning and big data have become an important direction for academic research and business needs. However, current research shows that even mainstream deep neural networks, trained with millions of samples, are easily affected by differences in domain data distribution, resulting in a sharp decline in learning performance. [0003] Therefore, people pay more and more attention to deep multi-task learning. Heterogeneity is a key property of big data, and it is ubiquitous and diverse, such as domain heterogeneity. In transfer learning and multi-task learning, data samples from d...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to overcome the above-mentioned defects in the prior art, provide a multi-task learning method and system based on feature and sample confrontation symbiosis, solve the problem of domain distribution differences and small samples, and greatly improve the generalization of machine learning systems performance

Images