Incremental Learning Based On Convolutional Neural Network Classifier

With the rapid development of information technology in recent years,data generated and acquired by people are increasing exponentially.It has become a crucial scientific research that how to mine and utilize these massive data.Machine learning is a new technology that uses computer to imitate human brain's learning and mining potential data characteristics.In reality,many data are obtained uninterrupted,which is inconsistent with the traditional batch machine learning.This issue should be solved by an incremental learning method,which can learn new knowledge from new data without interruption,and retain knowledge that has already been learned.Firstly,in this paper,a new pre-defined evenly-distributed class centroids loss function(PEDCC-Loss)is proposed for convolutional neural networks,which explicitly encourages intra-class compactness and inter-class separability between learned features,and achieves a good recognition rate in both image classification and face recognition.Secondly,based on the loss function,an incremental classifier based on SVDD is proposed.This algorithm uses convolutional neural network to extract features,and then combines the improved loss function to improve the feature distribution.Then it uses Support Vector Data Description(SVDD)to construct a single class of hyperspheres for image features,and achieves class-incremental classifier through the increment of hyperspheres.This algorithm can construct hyperspheres as classifiers to classify samples without providing negative samples;each hypersphere is relatively independent and will not affect each other when classifying,which greatly increases the robustness of the incremental learning system.Finally,an incremental network algorithm is proposed,which uses multiple convolutional neural networks to train their respective categories independently by PEDCC-Loss,and then combines multiple networks to extract features to classify the whole network,so as to realize incremental learning of the network.This algorithm integrates multiple networks,each network trains different data independently and does not interfere with each other,thus realizing incremental learning in real sense.In order to improve the final classification performance of the incremental learning algorithm,all the processes of the algorithm are optimized.In this paper,convolutional neural network is used to learn and express image features.In order to improve the quality of extracted features,image features are optimized by improving the loss function.In the incremental process of classifier,SVDD chooses appropriate penalty coefficients and kernel functions.In the incremental process of network,PEDCC-Loss is used to train the network independently,so that each classes can be trained.