Multimodal representation learning with neural networks

Abstract

Abstract: Representation learning methods have received a lot of attention by researchers and practitioners because of their successful application to complex problems in areas such as computer vision, speech recognition and text processing [1]. Many of these promising results are due to the development of methods to automatically learn the representation of complex objects directly from large amounts of sample data [2]. These efforts have concentrated on data involving one type of information (images, text, speech, etc.), despite data being naturally multimodal. Multimodality refers to the fact that the same real-world concept can be described by different views or data types. Addressing multimodal automatic analysis faces three main challenges: feature learning and extraction, modeling of relationships between data modalities and scalability to large multimodal collections [3, 4]. This research considers the problem of leveraging multiple sources of information or data modalities in neural networks. It defines a novel model called gated multimodal unit (GMU), designed as an internal unit in a neural network architecture whose purpose is to find an intermediate representation based on a combination of data from different modalities. The GMU learns to decide how modalities influence the activation of the unit using multiplicative gates. The GMU can be used as a building block for different kinds of neural networks and can be seen as a form of intermediate fusion. The model was evaluated on four supervised learning tasks in conjunction with fully-connected and convolutional neural networks. We compare the GMU with other early and late fusion methods, outperforming classification scores in the evaluated datasets. Strategies to understand how the model gives importance to each input were also explored. By measuring correlation between gate activations and predictions, we were able to associate modalities with classes. It was found that some classes were more correlated with some particular modality. Interesting findings in genre prediction show, for instance, that the model associates the visual information with animation movies while textual information is more associated with drama or romance movies. During the development of this project, three new benchmark datasets were built and publicly released. The BCDR-F03 dataset which contains 736 mammography images and serves as benchmark for mass lesion classification. The MM-IMDb dataset containing around 27000 movie plots, poster along with 50 metadata annotations and that motivates new research in multimodal analysis. And the Goodreads dataset, a collection of 1000 books that encourages the research on success prediction based on the book content. This research also facilitates reproducibility of the present work by releasing source code implementation of the proposed methods.