Research On Deep Transfer Active Learning Method

Deep learning has achieved significant results in many fields,while training a deep model from scratch requires a lot of labeling costs and training time.This paper alleviates this problem by a deep transfer and active learning(Deep TLAL)method.A dynamic sample selection strategy for active learning that strongly coupled with transfer learning is proposed.With an initial model transferred from source task,it selects objective task samples mostly contributable to improve model performance by using a dynamically weighted combination of two measures,difference from source domain and target domain uncertainty.A new term,information extraction ratio(IER),is proposed and concretely defined in the specific case,and a IER-based batch training strategy and a T & N batch training strategy that blends IER-based and normal batch training strategies are proposed for dealing with model training process.The proposed methods are tested in two cross-dataset transfer learning experiments,the results show that in comparison with base-line methods,Deep TLAL can achieve better performance and effectively reduce labeling cost,and the IER related training strategies could optimize the distribution of computing resources over the active learning process,that is making the model learn more times from samples in the early phrases and learn less times in the later and end phrases.