Continuous learning method and system based on generative model and meta-learning optimization method

Abstract

The invention discloses a continuous learning method and system based on a generative model and a meta-learning optimization method, wherein the method includes the following steps: establishing a deep learning model and a generative model; when multiple new tasks are received, using the meta-learning optimization method Train the generative model to generate simulated data; input the simulated data and the input data of multiple new tasks into the deep learning model for training at the same time, so that the trained deep learning model can perform different tasks according to any input data. This method combines the optimization method of generative model and meta-learning to solve the machine learning problem in the continuous learning scenario, and also solves the problem of how to transfer the knowledge learned by the model on the old task to the learning of the new task when multiple tasks arrive successively over time.

Description

technical field [0001] The invention relates to the technical field of machine continuous learning, in particular to a continuous learning method and system based on a generative model and a meta-learning optimization method. Background technique [0002] When the statistical machine learning system and deep learning system are put into use, the distribution of the input data and output data learned by the model is required to be consistent with the distribution to be tested, and the distribution of input data and output data over time should also be considered. . In a single-task machine learning scenario, the feasibility and effectiveness of statistical machine learning models or deep neural network models depends heavily on an assumption: the distribution of input and output data does not change significantly over time; The resulting patterns and complex relationships are poorly represented or even completely unusable. But in the real world, such assumptions rarely hold...

AI Technical Summary

Problems solved by technology

The disadvantage of the existing technology is that on the one hand, by constraining the change degree of model parameters, it will affect the learning effect of new tasks, and on the other hand, the effect of solving the forgetting problem is not good; by adding memory storage units, additional storage space will be added. When the storage space When limited, the model does not perform well in solving the forgetting problem

Images