Research On Deep Multi-instance Learning Models And Algorithms

In the field of machine learning,multi-instance learning(MIL),as a weakly supervised learning paradigm,has been widely applied in computer vision,text classification,medical image analysis,audio time detection and other fields.In recent years,with the rapid development of deep learning technology,deep model combined with neural network has become an important research focus in the field of multiinstance learning.Deep supervision is effective for learning the high-level instance representations,and deep models achieve better classification performance than the traditional shallow MIL method.However,on the one hand,existing MIL neural networks usually regard the instances in bags as independent and identically distributed,which ignores the fact that important structural information exists among the instances.On the other hand,many methods encode instance characteristics in scalar-form when constructing bag representations,which may not be sufficient to effectively retain the attribute information of the instances in the bags.Therefore,how to fully capture inter-instance information and bag attribute information becomes an important research direction of deep MIL.Based on the analysis of the research status of deep MIL algorithms,this paper conducts in-depth research on the deep MIL model and algorithm aiming at the above two problems,and obtains the following research results:1.A deep multi-instance learning model and algorithm based on self-attention.First,to solve the problem that existing algorithms rely on the assumption of independent and identical distribution of instances while ignoring important structural information between instances,this chapter presents a deep multi-instance neural network based on self-attention mechanism(SA-MINN),which integrates the global dependence between the instances with self-attention mechanism and learns very discriminative and different features for target and non-target instances within a bag.Then,to reduce the computational complexity,a self-attention framework based on inducing point(ISA-MINN)is proposed,which reduces the calculation time of selfattention from quadratic to linear without sacrificing performance.Both models are verified by experiments: the proposed methods are not only suitable for standard MIL assummption problems,but also has good classification ability under the condition of the assummption based on threshold and counting.Moreover,the proposed methods are especially suitable for medical image analysis tasks since it could aggregate multilayer dependency information between image regions.2.A deep multi-example learning model and algorithm based on nested bags.To solve the problem that the existing methods are not efficient in encoding representation of bags,a deep multi-multi-instance neural network(MMI-Net)is proposed by converting multi-instance representation into multi-multi-instance representation and combining multi-multi-instance learning with multi-instance learning.This method encodes bags of instances into bags-of-bags and extract instance features in vector form with convolutional networks.Therefore,the proposed method models multiple embeddings(channels)for each instance in the bag,capturing the attributes of the input instances from different aspects.MMI-Net then constructs a high-quality nested bag representation with the most discriminating instance features aggregated by the bag-Layers.Experimental results show that MMINet can learn transformation functions flexibly and adaptively according to different tasks and data.Compared with other MIL methods,MMI-Net is more effective in encoding instance features,which is very beneficial for processing large datasets.The first method captures the dependencies between instances with self-attention mechanism,the second method capture multi-aspect feature of instances in MMIL framework.The common point of these two methods is that neural networks are used to capture useful instance attributes and learn high-level bag representation to achieve good classification performance.Experimental results show that both methods achieve the state-of-the-art or competitive performance on a wide range of MIL datasets and have good interpretability.