Concerning that most of the current methods for Multi-Instance Multi-Label (MIML) problem do not consider how to represent features of objects in an even better way, a new MIML approach combined with Probabilistic Latent Semantic Analysis (PLSA) model and Neural Network (NN) was proposed based on topic model. The proposed algorithm learned the latent topic allocation of all the training examples by using the PLSA model. The above process was equivalent to the feature learning for getting a better feature expression. Then it utilized the latent topic allocation of each training example to train the neural network. When a test example was given, the proposed algorithm learned its latent topic distribution, then regarded the learned latent topic allocation of the test example as an input of the trained neural network to get the multiple labels of the test example. The experimental results on comparison with two classical algorithms based on decomposition strategy show that the proposed method has superior performance on two real-world MIML tasks.