Multiple Instance Learning (MIL) was originally proposed for drug activity prediction [1]. Now it has been widely applied to many domains and is an important problem in machine learning. Many multimedia data have the Multiple Instance (MI) structure; for example, a text article contains multiple paragraphs, an image can be divided into multiple local regions, and a gene expression data contains multiple genes. MIL is useful to processing and understanding MI data. Multiple instance learning is a kind of Weakly-Supervised Learning (WSL). Each sample is in the form of labeled bags, composed of a wide diversity of instances associated with input features. The aim of MIL, in a binary task, is to train a classifier to predict labels of testing bags, which is based on the assumption that a positive bag contains at least one positive instance, whereas a bag is negative if it is only constituted of negative instances. Thus, the crux of MIL is to deal with the ambiguity of the labels of the instances, especially in positive bags that have plenty of cases with different compositions. There are many algorithms have been proposed to solve the MIL problem. According to the survey by Amores [2], MIL algorithms are in three folds: instance-space paradigm, bag-space paradigm, and embedded-space paradigm. Instance-space paradigm learns the instance classifier and performs bag classification by aggregating the responses of instance-level classifier. Bag-space paradigm exploits bag relations and treats bag as a whole; in particular, it calculates bag-to-bag distance/similarity; then the nearest neighbor or Bayesian classifier carries out bag classification based on the distances/similarities. Embedded-space paradigm embeds a bag into a vocabulary-based feature space to obtain a compact representation for the bag, for example, a vector representation; then classical classifiers can be applied to solve the bag classification problem. Deep neural networks have been applied to solve many machine learning problems. For supervised learning, there are several kinds of neural networks. Deep Belief Networks (DBN) [3] use unsupervised pre-training and take a fixed length vector as input for feature learning, regression, and classification. Deep Convolutional Neural Networks (CNN) [4,5] take images as input and have dominated many computer vision problems. Deep Recurrent Neural Networks (RNN) [6] and Long Short Term Memory (LSTM) networks [7] take sequential data as input, such as text and speech, and are good at dealing with sequence prediction problems. Usually, training these deep networks requires a huge number of fully labeled data, that is, each training sample/instance needs a label.