Transfer Learning And Its Applications In Fine-grained Classification

Although great success has been achieved on tasks of vision and language via deep learning,the prerequisite of a large amount number of manually labeled samples largely limits its application in real-world tasks.To this end,we investigate how to train an efficient model for a given target task with few labeled training data or with few manual annotations.We take the task of fine-grained image classification as an example and study in the following three aspects.Firstly,we study how to discover the local features of an entire image in a weakly supervised manner.We propose a novel weakly supervised part-aware network with two parallel streams,where the classification and detection streams rank the part proposals for category classification and part detection,respectively.Semantic part proposals can be obtained on the detection stream with proper designs of model structure and loss functions.Then,we investigate how to transfer knowledge from the auxiliary source data to the target data in the task of domain adaptation.To promote the category-level feature alignment across domains,we introduce a domain symmetric network with the corresponding two-level domain confusion loss,where the category-level confusion loss improves over the domain-level one to promote the domain invariant features learning at a finer scale.Finally,we investigate how to apply the technique of transfer learning to the task of fine-grained image classification.We note that popular transfer learning methods,such as pre-training or multi-task learning,typically neglect the target task in the extraction of transferable information.To this end,we propose a regularized meta-learning objective based on the model structure of multi-tasks learning.We observe significant advantages by considering the target tasks via the regularized meta-learning objective when transferring knowledge from the auxiliary data to the target task.With careful experiments,we achieve the following three conclusions: 1.Excavating the local part knowledge can benefit the classification of fine-grained images.2.There are advantages to align the category-level feature distributions across domains,compared to the domain-level one,in domain adaptation.3.Taking the target task into the knowledge transfer process can lead to more valuable knowledge for the target task.