Research On Improvement Of Meta-Learning And Application In Ensemble Learning And Deep Learning

Few-shot learning is one research direction of meta-learning.Meta-learning means learn-to-learn,few-shot learning applies meta-learning methods to learn how to learn from a small amount of data.The work of this paper is carried out in two aspects:model design and data processing.We design a model controlling the learning process by limiting model parameters to low-latitude spaces.The establishment of this space is related to the current task.We call this method the DCN(Decoder Choice Network),which can establish the correlation between the model and the task with low time and space complexity.In terms of data processing,a new data augmentation method is proposed in this paper.The main goal of task augmentation is not to increase the number of data samples,but to increase the number of training tasks.The task augmentation is achieved by setting the new samples obtained from the image rotation as new data classification.The experiments are conducted based on the method with high performance,so we also applied the model combining ensemble learning and deep learning in our experiments.The experiments of DCN showed that the proposed model could get higher or equal well results with a small number of parameters.Experiments on task augmentation showed that the proposed data augmentation method could well reduce overfitting and improve the accuracy of prediction.Besides,this method is simple and inexpensive.Finally,this article conducted the research of applyinh of ensemble deep learning.It combines gradient boosting and 1D convolutional neural network for stock prediction.The experiments proved that ensemble deep learning could simultaneously learn the data fro different stock markets to improve the prediction accuracy.