DeepL earning Y annL eC un∗ Y osh uaB engio∗ G eoffrey H inton 深度学习 Y annL eC un∗ Y osh uaB engio∗ G eoffrey H inton A bstract Deepl earningal l owscomputationalmodel sth atare composed ofmul tipl e processing l ayers to l earn representations of data with mul tipl e l evel s of abstraction. T h ese meth odsh ave dramatical l y improved th e state- of- th e- artinspeech rec- ognition, visual objectrecognition, objectdetection and many oth er domains such as drug discovery and genomics. Deep l earning discovers intricate structure in l arge data sets by using th e backpropagation al gorith m to indicate h ow a mach ine sh oul d ch ange its internal parameters th at are used to compute th e representation in each l ayer from th e representation in th e previous l ayer. Deep convol utionalnets h ave brough tabout breakth rough sinprocessing images, video, speech and audio, wh ereasrecurrentnets h ave sh one l igh tonsequentialdatasuch astextand speech . 摘要 深度学习允许由多个处理层组成的计算模型学习具有多个抽象级别的数据表示。 这些方法极大地提升了语音识别、视觉目标识别、目标检测以及许多其他领域的 最新技术，例如药物发现和基因组学。深度学习通过使用反向传播算法来指示机 器应如何更新其内部参数（从上一层的表示形式计算每一层的表示形式），从而 发现大型数据集中的复杂结构。深层卷积网络在处理图像、视频、语音和音频方 面带来了突破，而递归网络则对诸如文本和语音之类的顺序数据有所启发。 正文 M ach ine- l earning tech nol ogy powers many aspects of modern society: from web search es to contentfil tering on socialnetworks to recommendations on e- commerce websites, and itis increasingl y presentin consumer products such as cameras and smartph ones. M ach ine- l earning systems are used to identify objects in images, transcribe speech into text, match newsitems, postsorproductswith users’ interests, and sel ectrel evantresul ts ofsearch . Increasingl y, th ese appl ications make use ofa cl assoftech niquescal l ed deepl earning. C onventionalmach ine- l earning tech niques were l imited in th eir abil ity to process naturaldata in th eir raw form. F or decades, constructing a pattern- recognition or mach ine- l earning system required carefulengineering and considerabl e domain expertise to designafeature extractorth attransformed th e raw data( such asth e pixel val ues ofan image) into a suitabl e internalrepresentation or feature vector from wh ich th e l earning subsystem, often a cl assifier, coul d detector cl assify patterns in th e input. R epresentationl earningisasetofmeth odsth atal l owsamach ine to be fed with raw data and to automatical l y discover th e representations needed for detection or cl assification. Deep- l earning meth ods are representation- l earning meth ods with mul tipl e l evel s of representation, obtained by composing simpl e but non- l inear modul es th ateach transform th e representation atone l evel( starting with th e raw input) into a representation at a h igh er, sl igh tl y more abstract l evel . W ith th e compositionofenough such transformations, very compl exfunctionscanbe l earned. F or cl assification tasks, h igh er l ayers ofrepresentation ampl ify aspects ofth e input th atare importantfordiscriminationand suppressirrel evantvariations. A nimage, for exampl e, comesinth e form ofanarray ofpixelval ues, and th e l earned featuresinth e firstl ayer ofrepresentation typical l y representth e presence or absence ofedges at particul ar orientations and l ocations in th e image. T h e second l ayer typical l y detects motifsby spottingparticul ar arrangementsofedges, regardl essofsmal lvariationsin th e edge positions. T h e th ird l ayermay assembl e motifsinto l argercombinationsth at correspond to partsoffamil iarobjects, and subsequentl ayerswoul d detectobjectsas combinations ofth ese parts. T h e key aspectofdeep l earning is th atth ese l ayers of features are notdesigned by h uman engineers: th ey are l earned from data using a general - purpose l earningprocedure. Deep l earning is making major advances in sol ving probl ems th ath ave resisted th e bestattemptsofth e artificialintel l igence community formany years. Ith asturned out to be very good atdiscovering intricate structures in h igh - dimensionaldata and is th erefore appl icabl e to many domains of science, business and government. In additionto beatingrecordsinimage recognitionand speech recognition, ith asbeaten oth er mach ine- l earning tech niques at predicting th e activity of potentialdrug mol ecul es, anal ysing particl e accel erator data, reconstructing brain circuits, and predicting th e effects of mutations in non- coding DN A on gene expression and disease. Perh aps more surprisingl y, deep l earning h as pro