¤Þ¥Î:
ì©«¥Ñ 41200 ©ó 2015-5-4 12:38 PM µoªí
·Q°Ý¤Ud **°ÝÃD..
¨Ì®aµo®iA.I ³Ì¤j¬Jªýê«Y«§??
¨S¤H²Î¤@¤j§½ÁÙ¨S¦³²©R©Êªº¤j²Î¤@²z½×
¤Þ¥Î:
ì©«¥Ñ 41200 ©ó 2015-5-4 12:38 PM µoªí
¥H¨Ì®a¬J¬ì§Þ, ¦³µL¥i¯àµo®i¨ì ironman ¤J±¬Jjarvis ¨t²Î¥X¾¤?
Jarvis¬O´X¤Q¥óªF¦èªº²V¦Xª«
¥Hvoice recognition¨ÓÁ¿, ²{¦³¬ì§Þ³£ÁÙ¥¼°÷¦n (§A¦³¸Õ¹L¥Î¤f¥hgoogle¶Ü)
¤Þ¥Î:
ì©«¥Ñ 41200 ©ó 2015-5-4 12:38 PM µoªí
·Q°Ý¤UÂI¼Ë±Nlearning system Åܨìlearn ¨ì¥ô¦ó³¥©O-.-?
·Q°Ý¤U«YA.I , ¹q¸£¦³µL¥i¯à¤ñÓinput ¦Poutput Ê\..
¤¤¶¡¬Jalgorithm ¥ÑÊ\¦a¦Û¤vcreate?
Artificial Neural Network
©Ò¦³parameters
¥ÑÊ\¦a¦Û¤vp
¦³Ó¤H¥Î¾ãÓYouTube
ªºvideo§@¬°input¥htrainÓlearning algorithm
µ²ªGÓnetwork¥i¥Hlearn¨ì¦p¦ó¤À¿ëþS©Mª¯ (without human specification)¤Þ¥Î:
ì©«¥Ñ 41200 ©ó 2015-5-4 12:38 PM µoªí
¦n¦üface recognition , ËݧA³£¸Ü¥i¥Hprogram ¥hp¬YD ³¥..
ÂI¼Ë¥i¥H¦n¦ü¤H¸£ËݱNµLprogram «×¬J³¥ ³£learn ¨ì?
EG §Ú¦a²ÓÓ«}¦³¨k¤k©OÓconcept ..
ËݧڦaÓ¸£´N·|¦Û¤vcreate ¤@Óalgorithm ¥h¤À...
«¢«¢, ¨ä¹ê¤°»ò³£¥i¥Hprogram
§A»{¬° "¤£¯àprogram¥hp" ¬O¦]¬°, §A¼È®É¤£ª¾¹D¥Î¤°»òparameters ¥hquantify¨ºÓ¨Æª«
§ä´M¤@ӼƾǪºfunction¥h´yz¤@¥ó¨Æª«,
¨ºÓ»â°ì¥sfeature extraction / machine learning
face recognition, ¨ä¹ê°µ¨ìÄê
¥i¯à§A¥Î/¾Ç¹Lªºalgorithm¤£°÷¦h
§A·í¤@Óinput face image¬OX, ¬OÓ3XNXMªºtensor
¦³2Ãþªºinput, ¨Ò¦p¨k¤Hªºface¥s type 1, ¤k¤Hªºface¥stype 2
¨ä¹ê§An§äªº¬O¤@Ófunction f, ³oÓf ¥i¥H§â³o2ÃþªºX, ²z¹L³oÓfunction¤§«á, Óvalue¦³¨º»ò»·¹j¨º»ò»·
§Y¬Odistance between f(X1) and f(X2) n·U¤j·U¦n:
find
§Úª¾§A¦³face recognitionªº¸gÅç
¨º»ò§Ú°Ý¤U§A
¦pªG§Ú¬ðµM¥s§A¥h°µ "dog-cat recognition"
§AÂI°µ
¦]¬°face-recognition¬O"text-book level" (¤w¸gÁ¿¨ìÄê), ©Ò¥H¦³"¸û§¹µ½"ªºknowledge
§A¥i¥H¥Î¤@¨Ç¨ä¥L¤H±`¥Îªº " f "
(§Y§A©ÒÁ¿ªº "program")
¦pªG§Ú¬ðµMµ¹§A ¤@¤d±i ¤û ©M ¤@¤d±i¦Ïªº¬Û¤ù , ¥s§A°µrecognition
§AÂI¸¨¤â???
¦pªG§A·P¨ì¦n °g¯í, ¤£ª¾¦p¦ó¤U¤â, ¨ºªí¥Ü§A¨S¦³ Feature Extraction ©M Dimension Reduction ªº»{ÃÑ
¦p²³æ, ¤£»Ýnª¾¹D¥ô¦ó¨Æªº±¡ªp¤U
§A±j¦æ§â¨ºÓ ´X¤d ´X¸UÓªºdata point ¶ë®Í¤JÓlearning algorithm, ³£¥i¥H
¦ý¬O¨º¼Ë°µ·|°ª¾÷·|·| "overfitting"
¦Ó¥B, ¤£¬O"·U¦hvariable´N·U¦hinformation content´N·U¦n"
¦³¤@¥ó¨Æ¥s
"Curse of Dimensionality"
·ídataªºvariable ¼Æ¶q·U¦h
learning ´N·UÃø (§Úassume§Aª¾¹D³oÓ"learning"¬O«ü machine learning³oÓfieldªºlearning, ¦Ó¤£¬O«ü"¤H¾Ç²ß")
©Ò¥H¤@©wn "°ºû" Dimension Reduction
¨Ò¦p
§â5-dimensionªºvector¥Î¤@Ó3-dimensionªºvector¥hrepresent
¦Ó2ªÌ¤§¶¡ªºdistancen·¥¤p
very interesting, recall ¤§«en§äªº "f"
¦Ó²{¦bn§äªº¬O g
¬°¤°»òn·ddimension reduction
¨ä¹êData Compression¬O¨ä¤¤¤@Óapplication
¦pªG§Anload Ó5-variableªº data, ¦pªGnload 100Ó, ¤@¦@n500Ómemory-space
¦pªG§A²{¦b¥u»Ýnload 3-variableªºdata, ¨º»ò²{¦b¥u»Ýn300Ómemory-space
·íµMdata compression¤£¬O¥Dnì¦]
¦]¬°²{®Éhardware¬ì§Þµo®i¦n
¤@¯ë¤Hªºmemory space ¤£·|¤£°÷ (°£«Dstore video of drama/ film, §_«h1TB space®Ú¥»¥Î¤£¤F¦h¤Ö)
ndimension reduction
¥Dn¬Ofor learning algorithm
¨Ò¦p§An§âÓ5-dimensional vector ÅÜ°µ1-dimension
³Ìsimpleªº¤èªk´N¬O Linear Regression
§Y¬O»¡, ¦pªG§Aª¾¹D ¨º¨Ç unknown a, §A´Nª¾¹D¤F³oÓ Âà´«ªºÃö«Y
Machine Learning´N¬On¥h"learn"³o¨Ça
¨º»ò¤°»ò¬ONeural Network
´N¬Omulti- linear regression
§A·íinput¬Ox, output ¬Oy
¦pªG¥Îlinear model, ¨º»ò layer-i , node -j ªº output ´N¬O
§Y¬O»¡ nlearnªº¬OÓmatrix
¨º»ò´N·|¦³´X¦ÊÓ´X¤dÓ unknown an¥hlearn
¦pªGdimension¦A¤j, number of unknown´N·|¶W¤j
·ínumber of unknowns ·U¤j, number of training samples´Nn §ó¥[¤j
Curse of Dimensionality «ü, ·ídimension¤j¹L¤@©wÈ, number of training sample´Nn °ò¥»¤Wn µLÓ
§Y¬Olearning ¬Oimpossible
¦ÓFeature Extraction, Dimension reduction´Nnn¤Æ¸Ñ³oÓ°ÝÃD
©Ò¥H, ¨ä¹êmachine learning, ÁöµM¨S¨º»ò²³æ
¤£¹L¥i¥H¥Î¤@±ørule generalize
¦pªG¬Odimension reduction, q<p
¦pªG¬Ofeature extraction , ´N¤°»òq³£¥i¥H