Sequence-to-sequence learning

Jeff Dean surveys how Google uses deep learning across products, research, and scientific applications, and explains how more data and compute have made neural networks effective on a growing range of problems. He describes TensorFlow’s goals and adoption, then gives examples including image understanding, medical diagnosis, robotics, scientific simulation, translation, and automated model design. The talk also covers TPUs and other systems work aimed at speeding up training and experimentation. In the Q&A, Dean discusses data efficiency, interpretability, reasoning, and where machine learning researchers can contribute.