Multi-instance Learning Using The Neural Network

Of late,deep neural network has achieved greate success.However,it requires a large number of data with accurate human-annotation,so weakly supervised learning is an important problem in machine learning and many other domains.Multi-instance learning(MIL)as a typical weakly supervised learning method,is effecitve for many applicatioins in computer vision,biometrics,natural language processing,and so on.In this article,we make a further study on multi-instance learning,combines it with popular neural network,and propose three multi-instance learning algorithms using neural network:(1)we propose the first embedded-space multi-instance neural network.Via the MIL pooling,it convert instance representations to the bag representation,and can be optimized the process of representation learning and bag classification in end-to-end manner.In addition,it is integrated with two popular deep learning tricks(deep supervision and residual connections)(2)we propose multi-instance neuarl network based on bag similarities.It is the first work that combines similarity learning with multi-instance learning.And it calculates the similarity between bags through the Hausdorff pooling to mine rich contextual information.Considering the characteristic of MIL and the complexity of model,we design a decoupled training scheme for effective learning.(3)we propose a multi-instance nerual network with dynamic pooling.It iteratively updates the instance contribution to its bag so as to adaptively select key instance and model contextual information.The dynamic pooling keep the permutation invariance of MIL and interpret the instance-to-bag relationship.To validate the above methods,we perform sufficient experiments on many MIL datasets.Results show the effectiveness of multi-instance learning with neural network.