23
Quoc V. Le Deep Learning: verview and Latest Result joint work with many Google & Stanford collaborators

Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

  • Upload
    mlconf

  • View
    1.317

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Deep Learning: Overview and Latest Results

joint work with many Google & Stanford collaborators

Page 2: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

tensorflow.org

Page 3: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Coates et al 2009

Page 4: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Feature Engineering

Handcrafted Features

Classifier

X X’

“Face”Y

Page 5: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Why Deep Learning?

x

Corner Detectors

Edge Detectors

Raw Image

Cat, Face Detectors

Page 6: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Model Partitioning

Machine 4

Machine 1 Machine 2

Machine 3

Page 7: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

GoogleBrain infrastructure

ParameterServer

ModelWorkers

DataShards

p’ = p + ∆p

∆p ∆p ∆p

Le et al, 2011, Dean et al 2011

Page 8: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Why scaling up?

Amount of data

Accuracy

Deep Learning

Traditional Machine Learning

Page 9: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

GoogleBrain teaches itself

Watch YouTube for 3 days (10M frames)

Run on 16,000 machines1 billion synapses

Le et al 2011

Page 10: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

GoogleBrain teaches itself

Watch YouTube for 3 days (10M frames)

Run on 16,000 machines1 billion synapses

“Cat” Le et al 2011

Page 11: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Page 12: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Applications

- Understanding Images

- Understanding Speech

- Understanding Texts

Page 13: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Image Recognition

28% vs. 5%Error rate:

Achieves superhuman performance

Now used in Google+ Photosearch and StreetView

LeCun et al 1998, Krizhevsky et al 2012

Page 14: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Speech Recognition

Voice Search

23% vs. 8%Error rate:

Hinton et al, 2011

Page 15: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Applications

- Understanding Images

- Understanding Speech

- Understanding Texts

Page 16: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Mapping sequence to sequence

A

W X Y Z <EOS>

B C <EOS> W X Y

W

Z

WX XWY YXWZ YXW<EOS>

Page 17: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Sample ConversationsDescribe your problem: i am having issues accessingvpn

Machine: hiHuman: helloMachine: could you please let me know what are the operating systems you are using ?Human: i am using linuxMachine: ok it was in the ssl vpn on which you stated it was brokenMachine: please check if you are logged in to corp networkMachine: what is the error that you are running pleaseHuman: i am seeing an error related to vpnMachine: what is the error message that you are getting when connecting to vpn using network connect ?Human: connection refused or something like thatMachine: may i know the version of network connect you connect ?

Page 18: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Sample ConversationsHuman: i am not sure i know thatMachine: is the network connect prompting for any user certificateMachine: ?Human: yesMachine: are you at home or at work at the moment ?Machine: office ?Human: i am at homeMachine: try thisMachine: goto <URL> page and enter your name and passwordHuman: doneMachine: try logging in with <NAME> and check if its working fine nowHuman: yes , now it works !Machine: greatMachine: anything else that i can helpHuman: no , that was very usefulMachine: you are other than <NAME> : )Machine: have a nice night

Page 19: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Sample Conversations

Human: what is the purpose of life ?Machine: to serve the greater good .Human: what is the purpose of living ?Machine: to live forever .Human: what is the purpose of existence ?Machine: to find out what happens when we get to the planet earth .Human: where are you now ?Machine: i ’m in the middle of nowhere .Human: what is the purpose of dying ?Machine: to have a life .Human: what is the purpose of being intelligent ?Machine: to find out what it is .Human: what is the purpose of emotions ?Machine: i don ’t know .

Page 20: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Applications

- Understanding Images

- Understanding Speech

- Understanding TextsImage to Texts

Page 21: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Generating captions for images

A person riding a motorcycle on a dirt road

A group of young people playing a game of frisbee

Two dogs play in the grass

Vinyals, et al, 2015

Page 22: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Deep Learning tutorials & software

Deep Learning tutorials are available on my website:

Part 1 (Basic): http://ai.stanford.edu/~quocle/tutorial1.pdf

Part 2 (Advanced): http://ai.stanford.edu/~quocle/tutorial2.pdf

Popular open source software for Deep Learning: CAFFE https://github.com/BLVC/caffe

Page 23: Quoc Le, Software Engineer, Google at MLconf SF - 11/13/15

Quoc V. Le

Thanks

Jeff Dean Oriol VinyalsIlya SutskeverGreg Corrado

Matthieu DevinAndrew Ng Rajat MongaNavdeep Jaitly

+GoogleBrain team, Tomas Mikolov, Thang Luong, William Chan, Wojciech Zaremba