基于典型样本的卷积神经网络增量学习研究

Incremental learning in convolutional neural network based on typical samples

基于深度学习的卷积神经网络,提出了一种选取典型样本与新增样本结合学习的增量学习方法,针对典型样本的选取方法和选择数量进行了简单的研究,发现新训练的样本总数与旧样本总数大致相当即可取得良好的效果,并且在分类MNIST手写数字中得到了印证。进一步研究了对旧样本中无法很好识别或识别错误的典型样本模式的处理方法,提出了结合典型样本和新样本训练的增量网络的概念,通过自适应的错分类样本增强实现了识别率的提升。

Based on convolutional neural network of deep learning,this paper presents an incremental learning method that typical samples and incremental network are chosen to combine with new samples to learn.We preliminarily research how and how many typical samples we should choose,and find that as long as the number of new samples almost reach the number of the old samples,it may achieve good results,and it has been confirmed in the classification of MNIST handwritten numbers.Furthermore,we research the treatment method for the old samples to form typical samples,and put forward the concept of incremental network trained with typical samples and new samples.The improvement of the recognition rate is achieved by the adaptive misclassified sample enhancement.