Resumo: In recent years, deep learning methods have been responsible for astonishing breakthroughs in computer vision, speech recognition, natural language processing, and robotics—among other applications. In this talk, I give an introduction to the topic of deep learning, illustrating with examples one of the greatest benefits of deep learning methods: the ability to learn directly from raw unstructured data (such as images and text) given large amounts of labeled data and computational power. I will show that a side effect of this learning process is that we are now able to use the internal representations learned by the model to map symbolic concepts into useful numerical representations (also known as “embeddings”) that can be processed by other algorithms. Finally, I will talk about generative adversarial networks (GANs), a more recent deep learning architecture that allows the generation of realistic synthetic data when trained on a set of real data and show an application developed by our team of researchers at IBM Research Brazil that uses GANs to generate realistic seismic data from geoscientist’s sketches .