Research On Key Algorithms Of Recurrent Attention For Weakly Supervised Learning

In recent years,deep learning is a new research hotspot in the field of artificial intelligence.Since deep learning is data-driven method,labeling and tagging data is usually done manually,which requires a lot of manpower and time.In addition,the traditional deep learning model encodes the entire input and is unable to find key parts of the input data.So how to train the deep learning model for weakly labeled data,and to find the key parts of the input becomes a problem worth studying.The Attention mechanism stems from the study of human vision.Due to the bottleneck of information processing,humans selectively focus on all the information while ignoring other visible information.In deep learning,the attention mechanism is usually composed of a neural network,which can assign different weights to different parts of the input or directly input some parts of the input.This feature makes the attention mechanism suitable for solving weak supervised learning problems.For the long document data and image data of weakly labeled data,two weak supervised classification algorithms are proposed in this paper,and the following aspects are completed:(1)This paper reviews the attention mechanism and various classical models based on the attention mechanism,and explains the source of the work of this paper.Secondly,the principle and process of the classical algorithm REINFORCE in reinforcement learning are introduced.(2)Aiming at the problem that the traditional text classification models use full-text information and only applies to short documents.This paper proposes a long document classification algorithm for weakly labeled data.The algorithm only needs to observe several groups of local keywords in a long document to classify the document.Firstly,the feasibility of the method is proved in principle,and then how to solve the problem of long document classification is explained step by step.Finally,it is proved by experiments that the method can correctly classify long documents and output key sentences in long documents by using only a small amount of text information.(3)For the weakly labeled image data,this paper proposes a fine-grained classification algorithm for images called Smart-zoom.Based on the attention mechanism and combined with deep reinforcement learning,the algorithm can imitate the human visual mode and gradually focus on the key positions of the image to improve the accuracy of classification.This paper theoretically proves the feasibility of Smart-zoom and elaborates the model workflow.Secondly,the two weakly labeled data used in the experiment is introduced.Finally,it is proved by experiments that the method can correctly identify the category of the image in the case of weak tags.