The Effect of Task Similarity on Deep Transfer Learning

Abstract

In recent years, with deep learning achieving a great success, deep transfer learning gradually becomes a new issue. Fine-tuning as a simple transfer learning method can be used to help train deep network and improve the performance of network. In our paper, we use two fine-tuning strategies on deep convolutional neural network and compare their results. There are many influencing factors, such as the depth and width of the network, the amount of data, the similarity of the source and target domain, and so on. Then we keep the network structure and other related factors consistent and use the fine fine-tuning strategy to find the effect of cross-domain factor and similarity of task. Specifically, we use source network and target test data to calculate the similarity. The results of experiments show that when we use fine-tune strategy, using different dataset in source and target domain would affect the target task a lot. Besides the similarity of tasks has direction, and to some extent the similarity would reflect the increment of performance of target task when the source and target task use the same dataset.