| Summary: | 碩士 === 國立中央大學 === 資訊工程學系 === 107 === In recent years, deep neural networks (DNN) have demonstrated their powerful learning ability in image recognition. Such success result requires a large-scale data set as training data. However, acquiring data label in real-life needs high cost. This dilemma condition makes the study of transfer learning.
In order to solve the problem of insufficient training data, transfer learning assumes that the training and test dataset are independent and identically distributed. Its purpose is to transfer knowledge from the source domain to the target domain. Thus, even if target task data set has only a small amount of label data or no label data, deep neural network still can be used to train and learn through existing label data.
Nowadays, adversarial-based deep transfer learning has been gradually applied in the domain adaption task. It uses a extra domain discriminator to learn domain-invariant features by minimizing the loss of domain discriminator and utilizing reversed gradient during the backpropagation process for aligning distributions.
The architecture proposed in this paper based on concept of collaborative and adversarial network(CAN),which is the first model using all of blocks features with high accuracy. It use all blocks with domain discriminators and let them learn domain feature except last one learning for domain-invariant feature by a reversal gradient layer. Our architecture also use all block features to advance transfer effect by weighted dense connection with a gradient reversal layer. In addition, our architecture add a classifier in domain discriminator, in order to make network retain classification capabilities when learning domain invariant features.
In the experiment, our architecture demonstrates better transfer ability in three transfer tasks of Office dataset. The accuracy of our proposed architecture in three transfer tasks are all about 0.3~0.5 higher than CAN.
|
|---|
