Research On Deep Multiple-instance Algorithms Based On Random Subspace

With the development of artificial intelligence and machine learning,deep learning has become known as one of the most popular areas of scientific research today,driving huge advances in computer vision and machine learning.The data required to build deep learning models often needs to be richly and accurately labelled,so weakly supervised learning,which requires only little supervision,has received attention from researchers.Multi-instance learning,which is weakly supervised learning,has been applied in computer vision,text classification,medical image analysis,and has become an important machine learning paradigm.Through the analysis of existing multi-instance learning algorithms,this paper combines the random subspace method with MI-Net algorithm and the construction of nested instance packages,respectively,to investigate deep multi-instance networks based on random subspace in depth,and the results of the study are as follows.1)A deep multi-instance learning method based on random subspace instance selection is proposed.The approach improves on the MI-Net,a multi-instance network based on embedding spaces,from a new perspective.First,use the fully-connected layer to obtain features of the bag that exemplify the inclusion of higher semantics;then,the positive score and instance selection probability are calculated by the random subspace instance selection module,and the instances are selected in this way to capture the relationships in the data and reduce the ambiguity of the instance labels;finally,the bag features are obtained by aggregating the instance features.The experimental results show that the proposed method achieves a better accuracy rate,up to about 3% higher than the previous method,and is feasible and effective.2)A deep multi-instance learning approach based on random subspaces and nested bags is proposed.In order to capture the structure of the bag and to obtain a reasonable and easy inference process for the data labels,this paper proposes to construct nested bags using random subspaces to transform multi-instance learning into a deeper learning framework.First,instance bags are transformed into nested bags by the random subspace method,and then,instance and sub-bag features are aggregated separately to obtain high quality representations of sub-bags and top-bags.The experimental results demonstrate the effectiveness of the conversion from multiinstance to nested multi-instance,improving the accuracy by up to 4% over existing methods,yielding better classification accuracy of bags and improving the interpretability of the algorithm.