Neural network continuous learning method and device based on task domain knowledge transfer

Abstract

This disclosure proposes a neural network continuous learning method, device, electronic device and medium based on task domain knowledge transfer. The method includes: extracting the features of the input image of the current task; classifying the input image of the current task according to the features; wherein, When the classification module corresponding to the current task can identify the features, use the classification module corresponding to the current task to classify the input image, otherwise, perform the following operations: use the first feature transfer module to transfer the feature to the first feature, use the classification module corresponding to the first task to classify the first feature; if the classification module corresponding to the first task cannot complete the classification of the current task, use the second feature migration module to migrate the first feature to the second For the second feature related to the task, use the classification module corresponding to the second task to classify the second feature; and so on until the classification result of the input image of the current task is obtained.

Description

technical field [0001] The present disclosure relates to the fields of deep neural networks and computer vision, and in particular to a continuous learning method and device for neural networks based on task domain knowledge transfer. Background technique [0002] Neural network-based continuous learning (Continual Learning), also known as Lifelong Learning (Lifelong Learning), Incremental Learning (Incremental Learning) and Sequential Learning (Sequential Learning), that is, learning continuous task capacity. Continuous learning has two key points. One is how to use the experience of previous tasks to complete the current task faster and better. The other is that when learning the current task, the tasks that have been learned before will not be forgotten. So far, this technical field still faces many challenges, the most important of which is the problem of catastrophic forgetting in neural networks. The so-called catastrophic forgetting means that after learning new kno...

AI Technical Summary

Problems solved by technology

The regularization-based method does not require additional storage space, that is, an additional regularization term is added to the loss function to consolidate previous knowledge while learning new data. This method may cause catastrophic forgetting due to insufficient punishment. , and as the amount of tasks increases, a fixed model capacity cannot learn an unlimited number of tasks

Images