Applications Of Mollifying Network For Optimizing Contrastive Loss Function In Face Recognition

In this paper,we mainly study the optimization problem of highly non-convex and non-linear loss function in depth learning.We introduce a mollifying network and apply it to the face recogni-tion technology which is very hot today.Mollifying network train the non-convex loss function by adding noise to the network,which inherite the continuation methods and annealing methods.First of all,we train a simple convex objective function,and then gradually increase the complexity of the objective function and until the final training,the original heighly non-convex and non-linear objective function will be trained.We first introduce the composition,including the concept of mollifiers and weak gradient,and training methods with noises of the mollifying network,and then introduce the contrasive loss function of the Siamese network for face recognition,which describes the similarity between photos with distance.In this paper,the mollifying network optimization method is used in the optimization of the loss function.we use face data from MUCT to do empirical analysis,training and testing the data with cross validation.We compare the results with the traditional method of the Siamese network in the learning effect and compare these methods in the aspects such as the loss function,the classification accuracy and the ROC curve.Finally,we find that the mollifying network in solving the optimization problem of the highly non-convex and non-linear loss function is more efficient and feasible with a higher classification accuracy.