Meta-learning method and related device

Abstract

The invention discloses a meta-learning method and a related device. According to the method and the device, the internal updating result of each round of N steps can be used for adjusting the attenuation coefficient of the first learning rate in the next round, so that the self-adaptive adjustment of the attenuation coefficient of the first learning rate is achieved. After the attenuation coefficient is updated, whether the preset training cut-off condition is met or not can be continuously judged, and when the preset training cut-off condition is met, training is stopped; and when the preset training cut-off condition is not satisfied, the updated attenuation coefficient is taken as the attenuation coefficient of the (k + 1) th round so as to execute the (k + 1) th round N step internal updating of the meta learner. According to the technical scheme, the attenuation coefficient of the first learning rate is adaptively adjusted according to the internal updating result of the element learner, the internal updating learning rate can be continuously reduced under the action of the attenuation coefficient, so that the internal updating effect of the element learner is improved, the internal updating process of the element learner is more efficient, and the process of seeking the optimal solution is more stable.

Description

technical field [0001] The present application relates to the field of computer technology, in particular to a meta-learning method and related devices. Background technique [0002] In machine learning, in order to solve technical problems in specific scenarios, a large amount of data in specific scenarios is usually used to train the model. However, when the scene changes, the model needs to be retrained. For humans, however, a child who grows up seeing pictures of many objects can, one day, see just a few pictures of dogs for the first time and be able to distinguish dogs from other objects quite well. Meta-learning means learning to learn, that is, learn to learn, which was born with this expectation of human beings' "learning ability". Meta-learning hopes to enable the model to acquire a "learning to learn" ability, so that it can quickly learn new tasks on the basis of acquiring existing "knowledge". For example, let AlphaGO quickly learn to play chess; let a cat pi...

AI Technical Summary

Problems solved by technology

In meta-learning, internal updates are performed with the same learning rate, which means that the weight adjustment range remains unchanged, which causes the network to repeatedly jump back and forth between different states, so it is difficult to stably seek the optimal solution, which also affects the meta-learner’s learning ability. Internal update effect

Images