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!