Image Classification Based On Curriculum Learning

Image classification is widely concerned because of its important application value in real-world.With the advent of the era of big data,deep neural networks have achieved excellent results in image classification tasks.However,most of deep model training methods treat all samples equally.Curriculum learning suggests that we should give different training priorities to different the samples according to their difficulty.In general,curriculum learning believes that deep models should learn simple samples in the early stage,and learn difficult samples in the later stage.There are two key points for the realization of curriculum learning.The first is how to measure the difficulty of samples.The second is how to implement the curriculum.To solve these challenges,we introduce curriculum learning into image classification tasks.The overall overview is as follows:In Chapter 2,we propose a more elaborate method to measure the difficulty of samples and implements a smoother curriculum.Most related works use the loss values of samples to measure their difficulties,which does not contain the relative position of samples and decision boundaries.We propose to use samples' predicted probabilities to measure their difficulties.This paper proposes a teacher-student training framework.In each training iteration,we first train the teacher network,and we use the predicted probability to estimate the difficulty of each sample.Based on the difficulties of samples,we implement our curriculum on the samples that train the student network.Many curriculum learning methods directly discarded some samples,which would waste a lot of data.We therefore propose to reweight the training samples to implement the curriculum.This approach allows the student network to focus more on simple samples in the early stages of training and focus on difficult samples in the later stages.We verified the effectiveness of our method in multiple data sets.In Chapter 3,we propose a framework for generating difficult samples around the boundary and using them to fine-tune the decision boundary.The samples around the decision boundary can help the classifier learn a better decision boundary.Therefore,we first use GANs to generate a large number of boundary samples that confuse the classifier,and then use co-distillation learning to help the classifier refine the decision boundaries.Many experiments on multiple datasets show the effectiveness of the proposed method.Finally,we analyze the similarities and differences of the above two methods and their advantages and disadvantages.