Learning Area @ ACM MM 2014

Deep Learning Area at ACM Multimedia 2014

Deep Learning is an emergent field of Machine Learning focusing on learning representations of data. Deep Learning has recently found success in a variety of domains, from computer vision to speech recognition, natural language processing, web search ranking, and even online advertising. Deep Learning’s power comes from learning rich representations of data that can be tuned for the task of interest. The ability of Deep Learning methods to capture the semantics of data is however limited by both the complexity of the models and the intrinsic richness of the input to the system. In particular, current methods only consider a single modality leading to an impoverished model of the world. Sensory data are inherently multimodal instead: images are often associated with text; videos contain both visual and audio signals; text is often related to social content from public media; etc. Considering cross-modality structure may yield a big leap forward in machine understanding of the world.

Learning from multimodal inputs is technically challenging because different modalities have different statistics and different kinds of representation. For instance, text is discrete and often represented by very large and sparse vectors, while images are represented by dense tensors that exhibit strong local correlations. Fortunately, Deep Learning has the promise to learn adaptive representations from the input, potentially bridging the gap between these different modalities.