[T BUFFET] 노정석 대표의 'How Computers Understand Humans'

How  computers    understand  human  

2013.06.12  노정석  

Brief  History  

•  16  year-­‐long  entrepreneurial  journey  •  4  @mes  of  founding  ventures  

– Each  of  which  went  IPO,  bankrupt,  M&A  with  Google,  …  

•  3  @mes  of  working  in  conglomerates  •  10+  companies  of  angel-­‐inves@ng  

– Many  cases  of  successful  exits  such  as  Ticket  Monster,  Dialoid,  PaprikaLab  …  

•  1  venture-­‐incuba@on  company  – Fast  Track  Asia  

Xenters SKC INTERVENTURE

Companies  I  founded

Companies  I  worked  for

Companies  I  invested

Went  public  in  2002

Went  bankrupt  2004

Acquired  by  Google  in  2008

Currently  in  love  with  

All  the  beginning  :  Feb.  1994  

Sun  sparc  sta@on  2  baram.kaist.ac.kr  [email protected]  KUS  

2  big  ques@ons  à  1  big  ques@on  

•  The  end  of  oil  era  – Sustainability  of  complex  society  

•  The  crea@on  of  human-­‐level  ar@ficial  intelligence  – When?  How?      

What  is  value  crea@on?  

Value  crea@on  is  ‘bringing  order  to  chaos’  itself.  

Source:  ‘Extropy’  by  Kevin  Kelly  

Evolu@on  is  all  about  organizing  informa@on  ‘beder’  

Adding  Neocortex  was  all  the  beginning.  

Evolu@on  of  compu@ng  

Ar@ficial  intelligence  will  be  a  new  epoch  for  evolu@on  

Neocortex  will  have  a  new  extender.  

Ray  calls  it  ‘Singularity’  

How  computers    understand  human  

2013.06.12  노정석  

Ques@on  #1  

이런 날이 올거라고 생각하시는 분 ?  

Ques@on  #2  

언제 즈음 나올 것 같나요 ?    1.  5년내  2.  10년내  3.  50년내  4.  인간의 신의 창조물이다. 그런날은 결코 오지 않는다.  

Answers  

#1.  Very  soon      (2029,  Ray  Kurzweil)  #2.  prac@cal  level  ?  In  5  to  10  years                    human  level?    In  20~30  years  

Human  vs.  Computer  

IBM  watson  

IBM  watson  

•  200m  pages  of  document  (4TB)  •  A  cluster  of  90  IBM  Power  750  Servers  

– 10Racks  – 2880  Power7  processor  cores  – 16  TB  of  RAM  

•  It  can  process  500  GB  of  data  in  a  second    – Equivalent  to  1M  books  

•  It  costs  3M  USD,  the  94th  fastest  supercomputer  

How  does  ‘it’  work?  

Siri  

Siri  

How  does  ‘it’  work?  

User  

Speech  Recogni@on  

Natural  Language  Understanding

Dialogue  Manager    

Natural  Language  Genera@on

Text-­‐to-­‐Speech  Synthesis

How  does  ‘it’  work?  

Con@nuous  Speech  Recogni@on

Dialogue  Management

Task  Comple@on

“국립중앙박물관으로 가는 길을 알려줘”

Recognized-­‐Speech  =    “국립중앙박물관으로 가는      길을 알려줘”

find_route  (      from=here,      to=“국립중앙박물관”  )

Confirm(        first_candidate=“국립중앙박물관”,      first_geocode=“용산구 용산동6가 168-­‐6”,      second_candidate=“국립국악박물관”,      second_geocode=“서초구 서초동 700”  )  

“서울 용산구에 있는 국립중앙박물관으로 가는 길을 원하십니까?

“응,  실시간 교통 정보 이용해서 경로 찾아줘”

Recognized-­‐Speech  =  “응,  실시간 교통 정보 이용해서 경로 찾아줘”

find_route(      from=here,      to=“국립중앙박물관”,          search_opIon=USE_RTTI  )  

start_navigaIon(      string=Default      opIon=USE_RTTI  )

“실시간 교통정보를 이용하여 길 안내를 시작합니다.

*  RTTI  :  Real-­‐Time  Traffic  Informa3on    

Acoustic Model

Language Model

Builder

Blog Twitter News

Crawler

Acoustic DB

Language Model

Acoustic Model Trainer

Decoder

Text Analysis (Grapheme-to-Phoneme)

Dictation

Really  simple^2    explana@on    

…    

Human  brain  

•  Neocortex    :  80%  of  brain  mass  •  Simple  homogeneous  circuit  structure  

– Brain  is  very  plas@c!  – Use  it  or  Lose  it  /  Fire  together,  wire  together  

•  300M  modules,    100  neurons  per  each  module  •  Each  module  is  one  padern  recognizer  •  Connec@on  maders,  100  trillion    

– Learning  makes  connec@ons  – A  lot  of  redundancy  

Padern  recogni@on  in  neocortex  

What  are  the  recently  solved  problems?  

•  Search  (Google  Knowledge  Graph)  •  Con@nuous  Speech  Recogni@on  •  Speech  Synthesis  •  Machine  Transla@on  •  Natural  Language  Understanding  

– Deep  Q&A  – Task  Comple@on  

•  Gene  predic@on  

What  has  changed  in  the  last  decade?  

•  All  the  theories  are  nearly  30~40  year-­‐old  already-­‐solved-­‐problems  mathema@cally,  only  the  prac@cal  implementa@on  started  working  recently.  

•  What  is  the  main  factor  that  enabled  this  *leap*?    

Where  we  are  now  

•  June  25,  2012  •  Lead  by  Andrew  NG  

–  Standford  professor  –  Coursera  founder  

•  1000  computers  with  16,000  processors  •  10m  200x200  s@ll  cuts  from  youtube  to  neural  networks  for  3  days  

•  More  than  1  billion  connec@ons  •  S@ll  long  way  to  go  for  complete  visual  cortex  simula@on.  Maybe  in  a  decade?  

Key  Takeaways  

•  Computer  is  not  just  aiding  tool  any  more,  it’s  becoming  intelligence.  – Most  of  white  collar  work  will  go  away.  – The  real  meaning  of  big  data  is  …  

 •  Do  not  step  away  with  fear,  embrace  it  more!  

– Think  like  computer  scien@st.  – You  can  hire  thousands  of  knowledge  workers  with  nearly  zero  price.  

Key  Takeaways  

•  Computer  is  not  just  aiding  tool  any  more,  it’s  becoming  intelligence.  – Most  of  white  collar  work  will  go  away.  – The  real  meaning  of  big  data  is  …  

 •  Do  not  step  away  with  fear,  embrace  it  more!  

– Think  like  computer  scien@st.  – You  can  hire  thousands  of  knowledge  workers  with  nearly  zero  cost.  

Cri@cal  intersec@on  right  ahead  

We’re  now  here,  *again*  

2  different  species,    that  were  originally  one,  human.  

Think  big!