Research And Application Of Neural Network In Few Shot Learning

Neural Network is a system that processes information by mimicking the structure and function of the human brain.It almost relies on a large number of labeled highquality samples.But actually,the human brain is better at recognizing different objects in a few samples.It is difficult to accumulate enough data or not easy to sample in many real situations,which limits the development and application of artificial intelligence.Therefore,how to make models learn fast learning skills has become a very important research direction.Noise has a big impact on the results in Few Shot Learning.The algorithm can easily distinguish key information and noise in big data,so it’ s easy to distinguish between useless information(detailed features of the same category)and useful information(key distinguishing features of different categories).In contrast,it is difficult to find such information in a few samples,because the sample size is too small.It is crucial to add some other methods to enhance modeling capabilities.Besides,Some changes in shooting angle or actual conditions also cause differences in images,which may lead to the data points of the same class to be still very discrete after mapping,and cannot be well aggregated.The main contents of this thesis are as follows:First,this thesis reviews the development of Few Shot Learning and introduces three methods based on metric learning.Although these methods have solved some problems in Few Shot Learning and proposed new research ideas,they have not solved the key problems of serious noise,insignificant feature extraction,and sensitivity to abnormal data.Second,in response to the above defects,the Projective Discriminant Network is presented.This thesis proposes a new embedding module in the shallow neural network,referring to the idea of discriminant analysis.It alleviates the problem of noise and makes the features of different categories more obvious.The mean of different categories is larger and the mean of the same category is smaller.Third,after discussing whether the space is linearly separable,this thesis proposes the Projective kernel Discriminant Network.The kernel function is introduced into the embedded module,which makes the embedded structure more flexible and changeable.The model tries to distinguish different categories in higher dimensions by projecting onto a higher latitude embedded public subspace.Forth,on the benchmark datasets MiniImageNet and CUB,the performance of the Projection Discrimination Network is verified and compared with the most advanced model.In fine-grained experiments,the Projection Discrimination Network has achieved high accuracy.The advantages and disadvantages of the two embeddings are compared at different levels.Visual display and exploratory analysis of the mapped data points show that the Projection Discriminant Network is reasonable and effective.Besides,on the benchmark datasets MiniImageNet and CUB,the performance of the Projection Kernel Discrimination Network is also verified.Experiments show that the Projection Kernel Discriminant Network has higher accuracy and is more suitable for different types of Subtasks.