Meta-learning-based domain increment method

Abstract

The invention discloses a meta-learning-based domain increment method. The method comprises the following steps: S1, constructing a pre-training model; s2, training an old model by using the pre-training model; s3, training a new model. According to the invention, randomly reserved 5% memory data and new data are mixed and finely tuned to train a new model, and a cross entropy loss function and aknowledge distillation loss function are combined to guide the learning of the new model, so that the new model learns the classification knowledge of the data in the new field while memorizing the classification knowledge in the old field, and the expenditure of data storage and training time is greatly reduced.

Description

technical field [0001] The invention relates to the field of computer technology, in particular to a field incremental method based on meta-learning. Background technique [0002] With the rise of deep learning, object classification methods based on convolutional neural networks have developed rapidly, and the recognition accuracy has been greatly improved. However, methods based on convolutional neural networks also have drawbacks: when the distribution of test image data is inconsistent with the distribution of training image data, such as changes in lighting, background, and pose, the accuracy of the model will decrease. Therefore, when new domain data appears, that is, data that is inconsistent with the original training data distribution, the model needs to be able to incrementally learn new domain classifications, that is, to learn new domain knowledge classifications while remembering the classifications of old domain data . [0003] At present, the most intuitive ...

AI Technical Summary

Problems solved by technology

However, the method based on the convolutional neural network also has defects: when the distribution of the test image data is inconsistent with the distribution of the training image data, such as changes in lighting, background, posture, etc., the accuracy of the model will drop

[0003] At present, the most intuitive domain incremental learning method is to continue training the model with data in new domains, but the accuracy of this method often cannot meet the requirements: if the training is insufficient, the accuracy of the data in the new domain is not high; The accuracy rate of data in the old field will decrease, and it is difficult to reconcile the two

However, if the old domain data and the new domain data are directly mixed to retrain the convolutional neural network, the data storage and training time will be huge, especially in practice as the new domain data increases, the overhead will increase.

Images