Multiple Instance Learning Algorithms Based On Deep Learning

Recently,weakly-supervised learning(WSL)has gained greater attention in the machine learning field since it significantly reduces the workload of manual annotation.As a typical weakly-supervised learning,Multi-instance learning(MIL)has been widely employed in different tasks,including object detection,semantic segmentation,medical diagnosis,scene classification,etc.With the wide application of deep learning technology,the multi-instance algorithm model based on deep neural network has achieved great success compared with the shallow learning algorithm model,and has become a research hotspot.MIL algorithms can be categorized into three groups,i.e.,instance-space algorithms,bag-space algorithms,and embedded-space algorithms.In this paper,from the perspective of capturing and highlighting important correlations between instances,reducing model complexity and reducing model parameters,the deep multi-instance models ARP-MINN and CSP-MINN are constructed based on the embedded-space algorithm paradigm and combined with deep neural networks.The main work and innovations of the paper are as follows:(1)A deep multi-instance learning algorithm based on adaptive recurrent pooling(ARP-MINN)is proposed.The algorithm adopts two components to model the context information within the bag,so as to solve the problem that the correlation between instances cannot be accurately captured in the corresponding scenario.Firstly,the multi-view global structural feature information from the whole bag through the self-attention mechanism in the feature extraction module.Secondly,the adaptive recurrent pooling operation aggregates the multi-view features after the feature extraction module to obtain the entire bag representation to further capture the contextual information inside the bag.In addition,to reduce the computational complexity of the traditional self-attention mechanism,the cross-normalization operation is introduced into the self-attention operation with a low-rank factorization to reduce the quadratic complexity caused by the softmax operator to a linear complexity.The algorithm has been experimentally validated to achieve competitive results on multi-instance benchmark datasets,especially achieving state-of-the-art performance with good interpretability in medical image classification scenarios.(2)A deep multi-instance learning algorithm based on clonal selection pooling(CSP-MINN)is proposed.In order to solve the problem of characterizing the relationship between instances,mining key instances within the bag,and the requirement of lightweight deep neural network,the algorithm proposes a trainable multi-instance pooling function based on the clonal selection algorithm,called clonal selection pooling,under the premise of introducing the multi-head attention to capture inter-instance correlations between instances.The pooling function optimizes the instance-level representation inside a bag by minimizing the bag-level losses,which can effectively obtain the relationship between instances and bags,and can reduce the amount of model parameters,speed up the convergence of deep neural networks,and make the model easy to train.Experimental results show that the algorithm achieves state-of-the-art performance on medical text diagnose dataset and obtains competitive results on multi-instance benchmark datasets.