微软邓力：驱动大数据人工智能多种应用的三类深度学习模式.pdf

for Wide-Ranging AI Applications with Big Data 驱动大数据人工智能多种应用的三类深度学习模式 Li Deng Chief Scienst of AI Microso (Research and ASG), Redmond, WA, USA Keynote at CAAI (中国人工智能大会); August 26, 2016 Thanks go to many colleagues at Microso and collaborang universies Definion Deep learning is a class of machine learning algorithms that •  use a cascade of many layers of nonlinear processing •  are part of the broader machine learning field of learning representaons of data facilitang end-to-end opmizaon •  learn mulple levels of representaons that correspond to different levels of abstracon •  …, … 2 Three Paradigms of Deep Learning •  Deep Supervised Learning –  Paired input-output big training data for predicon –  Paired output serves as “teacher” for corresponding input •  Deep Reinforcement Learning –  Very weak “teacher” in the form of rewards; i.e. feedbacks (oen distant ）from environments –  Q-learning computes “teaching signal” for training DNNs •  Deep Unsupervised Learning –  Unpaired input-output bigger training data for predicon - but no teacher/label per se for each input token –  Non-predicon tasks (clustering, dimensionality reducon, interpretaon understanding of data for transfer/multask learning …) 3 Two Types of Big Data for D