This document discusses deep learning and its applications to multimedia. It provides background on shallow models from the late 1980s and more recent developments in learning with structured data since 2000. Deep learning has seen significant progress since 2006 with improved neural networks. Convolutional neural networks achieved state-of-the-art results on the ImageNet challenge in 2012. Deep learning is also being applied successfully to speech recognition, natural language processing, and other domains. Restricted Boltzmann machines and deep belief networks are important building blocks for deep learning models. Layerwise pretraining and fine-tuning have led to improved classification results.