Research On Multimodal Representation Based On Heterogeneous Ensemble Learning

With the rapid increase of social capacity,human beings have entered the era of artificial intelligence of information explosion.E-commerce,medical treatment,transportation,logistics and other industries with a large volume of production of rich data,which led to the era of data revolution.Data and algorithm are the foundation of the information age led by computer,and also the foundation of machine learning and deep learning as the core technology of artificial intelligence.This paper starts from machine learning and deep learning,mainly aiming at heterogeneous integration learning in traditional machine learning and image description task represented by multimodal representation in deep learning.The main research are as follows:Stack generalization has the inherent high complexity and data leakage problem,and there are also stability problems for different data samples.In this paper,we propose a LSH-bag-dag-stacking(LBDS)algorithm.It maps training sets and test sets to hash buckets by using local sensitive hashing(LSH)algorithm.When one of them is full,the model is used to predict the training data and test data and their neighborhood,The base classifier is selected by using the stability and information entropy conditions to generate high-level data.Finally,the results of the prediction of high-level training are obtained by mixed voting and average method.The results of verification on several datasets show that LBDS has better stability and stronger generalization ability.Most of the current image description models are mediocre and ordinary,and they can not generate the description with the focus of the object of concern According to different user's knowledge experience.In this work,this paper proposes a subjective and heterogeneous attention(SCHA)structure to represent the user's "knowledge experience",and control the description content and the level of detail.The prior knowledge graph with subject consciousness in the SCHA contains a directed graph composed of three nodes:object,attribute and relationship.It is abstract from image and has no corresponding specific semantics.Through acquiring scene map information from different angles as prior knowledge,SCHA significantly improves the description level,and realizes better Accuracy and richness than the well-designed baseline model on MSCOCO and VisualGenome datasets.Finally,the simulation system of the verification algorithm is designed,the system runs stably,can realize the reasoning result of the algorithm model,and achieve the expected goal,which has certain guiding significance for the development and design of the simulation system of related algorithm.