Consider the problem of building high-level, multimodal features from only unlabeled data, we train model consisting of a sparse stacked denoising autoencoder network with max pooling, which can be used to extract high-level image feature, on a large dataset consisting of multimodal information, and a text treating processes. Our model joints the image feature and text feature as representation of one united movie. We find that these representation can be used in regression mission, predict movie's rating, and the model obtains better effect than unimodal representation.
展开▼
