Neural network model multiplexing method based on transfer learning

Abstract

The invention discloses a neural network model multiplexing method based on transfer learning. The method includes measuring the similarity of two distributions of the target domain and the source domain according to MMD (Maximum Mean Difference); judging whether the distribution of the target domain and the source domain is the same or not according to hypothesis testing; if the hypothesis test judges that the two distributions are the same, performing finetune adjustment on the deep neural network model, and realizing multiplexing of the deep neural network model.

Description

technical field [0001] The invention belongs to the field of neural network algorithm model reuse. The technology is based on migration learning, so that the neural network algorithm model can be reused in similar subjects for reference by users. Background technique [0002] As the amount of business data increases, enterprises often regard multi-faceted topic analysis as an independent project, that is, they need to re-acquire data, re-perform data preprocessing, and rebuild models every time they conduct topic analysis. When conducting multi-subject analysis, this approach often leads to problems such as poor reusability of algorithmic models among subjects. In practical applications, it is usually not necessary to train a neural network from scratch for a new task. Such an operation is very time-consuming. In particular, the training data cannot be as large as ImageNet, and a deep neural network with sufficient generalization ability can be trained. Even with such a l...

AI Technical Summary

Problems solved by technology

Even with so much training data, the cost of training from scratch is unaffordable

Images