Adapting RBF neural networks to multi-instance learning

In multi-instance learning, the training examples are bags composed of instances without labels, and the task is to predict the labels of unseen bags through analyzing the training bags with known labels. A bag is positive if it contains at least one positive instance, while it is negative if it contains no positive instance. In this paper, a neural network based multi-instance learning algorithm named RBF-MIP is presented, which is derived from the popular radial basis function (RBF) methods. Briefly, the first layer of an RBF-MIP neural network is composed of clusters of bags formed by merging training bags agglomeratively, where Hausdorff metric is utilized to measure distances between bags and between clusters. Weights of second layer of the RBF-MIP neural network are optimized by minimizing a sum-of-squares error function and worked out through singular value decomposition (SVD). Experiments on real-world multi-instance benchmark data, artificial multi-instance benchmark data and natural scene image database retrieval are carried out. The experimental results show that RBF-MIP is among the several best learning algorithms on multi-instance problems.