Domain adaptation method based on deep network and confrontation technology

Abstract

The invention discloses a domain adaptation method based on a deep network and a confrontation technology, and relates to deep learning, migration learning, domain adaptation, convolutional neural network, anti-network technologies. On the basis of fine-tuning Alexnet, we added two confrontation subnets to correct the differences between samples in different fields and to learn shareable featuresat the high-level layer. This method can effectively reduce the cost of manual marking in a big data environment, and has certain practical significance. The algorithm proposes innovation based on theupper bound of a new target risk error. The algorithm mainly includes an initialization phase and a network training phase. In the initialization phase, a new neuron layer is constructed based on thenew error upper bound, and the corresponding loss and regularization terms are added, and the network and a data set are initialized. In the training phase, an original hyperparameter is replaced bya probability threshold, and a plurality of iteration cycles are run according to the probability iterative SGD algorithm until a condition is met, and the training ends. The final trained network caneffectively replace the manual marking process to obtain more and more accurate marked samples.

Description

technical field [0001] The invention discloses a field adaptation algorithm based on deep network and confrontation technology. The algorithm involves convolutional neural network, deep learning, and machine learning optimization, and belongs to the field of artificial intelligence. In particular, it involves a new error based on formula derivation. The last session increases the multi-branch structure and loss function. After confronting in the subnetwork, the pre-trained AlexNet is fine-tuned. It involves a new combined confrontation network, which can effectively complete artificial intelligence. The domain adaptation (transfer learning) task of knowledge representation transfer between different domains. Background technique [0002] Dataset skew is a problem that cannot be ignored in the field of machine learning. The data set is a one-sided representation describing the objects in the real world. If a model with the same structure is trained on a data set describing t...

AI Technical Summary

Problems solved by technology

The domain adaptation method of the present invention can save a lot of human labeling work, and has wide practicability for solving the problem of scarcity of labeled data under big data

Images