A Continuous Learning Method Based on Gate Control Mechanism

Abstract

The invention discloses a continuous learning method based on a gate control mechanism. Firstly, based on a deep neural network model, a corresponding model is established for the current task; secondly, a new specific depth model is established according to a new task, and the new model is compared with the original task. Corresponding connections between the layers are established between the models, so that the features of the old and new models on the same layer can be converged into the new model; then, a mechanism based on gate control is built on the connections of the same layer to learn the feature relationship between the old and new tasks, and to filter out the old tasks. Useful information for the new task; finally, after building the above model, train the model on the new task data. The present invention is an effective, high-precision, and distributed continuous learning method under the current general-purpose task prone to catastrophic forgetting and difficult to transfer knowledge of other tasks.

Description

technical field [0001] The invention relates to a continuous learning method based on a gate control mechanism, which belongs to the field of artificial intelligence. Background technique [0002] In recent years, deep learning technology has made remarkable achievements in many fields. However, the current deep learning models are mainly designed for closed environments and for specific tasks. Although multi-task models can perform multiple tasks, they assume that the tasks are highly correlated and require that all data about the tasks be collected. To train the model under certain conditions, adding new data requires saving the previous data and training with the new data, which will lead to a sudden increase in the amount of calculation and take up a large amount of storage space. Facing a real open environment, it is often necessary to perform multiple tasks, requiring the agent to have the ability to gradually learn multiple tasks like humans, that is, continuous lear...

AI Technical Summary

Problems solved by technology

Catastrophic forgetting is mainly caused by the destruction of the parameter space of the previous task in the learning of the subsequent task. The distributed continuous learning model does not consider the relationship between the current task and the previous task, and simply relies on the transfer matrix to introduce the feature flow of the previous task into the current task. task network

If there is no intersection between task feature spaces, violent migration is not beneficial to the current task learning or even harmful to learning

Images