Method for overcoming catastrophic forgetting based on parameter importance

Abstract

The invention discloses a method for overcoming catastrophic forgetting based on parameter importance. The method comprises the following steps of: after training a first task, testing the performanceof a model by using test data of the first task, and then calculating the importance of each parameter in a network model to the task by using the training data of the first task by using the methodfor calculating the parameter importance provided by the invention; secondly, adding the method provided by the invention as a regular item into a loss function in the model, and testing the performance of the model by using the test data of the current task and all previous tasks after training is completed; calculating parameter importance by using training data of a new task according to the method provided by the invention, and accumulating the parameter importance with a previously calculated parameter importance matrix; and finally, repeating the above steps when a new task is entered totrain the new task. Experiments prove that the method provided by the invention can effectively alleviate the problem of catastrophic forgetting in the deep learning model.

Description

technical field [0001] The invention relates to a method for overcoming catastrophic forgetting based on parameter importance, belonging to the field of artificial intelligence. Background technique [0002] For long sequence tasks, humans can learn in a continuous manner, old, rarely used knowledge can be covered by new incoming knowledge, while important, often used knowledge will not be forgotten, so that it can be gradually Learn more knowledge. In the deep learning model, if a new task is trained on an already trained model, the model is prone to "catastrophic forgetting", that is, the model cannot maintain the original task after learning multiple tasks continuously. Performance, the main reason is that the parameter configuration of the new task will overwrite the original parameter configuration, resulting in the destruction of the original parameter space. The catastrophic forgetting problem seriously restricts the performance of the agent when dealing with long s...

AI Technical Summary

Problems solved by technology

[0007] The purpose of the present invention is to provide a method based on parameter importance to overcome catastrophic forgetting, which can effectively alleviate the problem of catastrophic forgetting in deep learning models, so that the model can retain the learning ability of previous tasks while learning new tasks

Images