11
13/08/15 1 A Short(er) Introduc0on To Deep Learning Dr. Brian Mac Namee University College Dublin

Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

Embed Size (px)

Citation preview

Page 1: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

1

A  Short(er)  Introduc0on  To  Deep  Learning  Dr.  Brian  Mac  Namee  University  College  Dublin  

Page 2: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

2

Deep Learning

Google Trends: http://www.google.com/trends/

2005 2007 2009 2011 2013 2015

Kaggle Digit Recogniser Contest https://www.kaggle.com/c/digit-recognizer

MNIST Dataset from Yan LeCun http://yann.lecun.com/exdb/mnist/index.html

Page 3: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

3

The standard approach to using machine learning to build a system to recognise these different digits is to first engineer a high-level respresentation

Percent filled: 0.37 Number of loops: 2 Direction 0: 0.2 Direction 1: 0.6 Direction 2: 0.1 Direction 3: 0.4

Percent filled: 0.11 Number of loops: 0 Direction 0: 0.1 Direction 1: 0.4 Direction 2: 0.0 Direction 3: 0.5

0 1

2

3

Percent filled: 0.29 Number of loops: 1 Direction 0: 0.2 Direction 1: 0.5 Direction 2: 0.4 Direction 3: 0.2

The standard approach to using machine learning to build a system to recognise these different digits is to first engineer a high-level respresentation

Percent filled: 0.37 Number of loops: 2 Direction 0: 0.2 Direction 1: 0.6 Direction 2: 0.1 Direction 3: 0.4

Percent filled: 0.11 Number of loops: 0 Direction 0: 0.1 Direction 1: 0.4 Direction 2: 0.0 Direction 3: 0.5

0 1

2

3

Percent filled: 0.29 Number of loops: 1 Direction 0: 0.2 Direction 1: 0.5 Direction 2: 0.4 Direction 3: 0.2

Using this reperesentation (6 features) we could train a decision tree that would manage to correctly

recognise about 8 out of every 10 digits

Page 4: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

4

Engineering  representa0ons,  is  one  of  the  most  important  and  >me  consuming  jobs  in  most  

predic>ve  analy>cs  projects,  and  needs  a  blend  of  technical  exper>se  and  domain  exper>se  

 

 Representa0on  learning  is  a  set  of  methods  that  allows  a  machine  to  be  fed  with  raw  data  and  to  

automa>cally  discover  the  representa>ons  needed  for  detec>on  or  classifica>on  

[LeCun  etal,  2014]  Deep Learning Yann LeCun, Yoshua Bengio & Geoffrey Hinton http://www.nature.com/nature/journal/v521/n7553/full/nature14539.html

Rosenbla='s  perceptron  from  1957  was  the  earliest  example  of  representa0on  learning,  and  the  first  neural  network  

Page 5: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

5

Each image is composed of 28 x 28 = 784 pixels each

containing a grayscale value between 0 and 255

Page 6: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

6

This low-level representation uses a vector of 784 features, each with values

between 0 and 255

Page 7: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

7

0

1 2

3

4 5

6 7

8

9

A simple representation learning

approach to the digit recognition problem could use a multi-

layer perceptron to make predictions using

the low-level representation

0

1 2

3

4 5

6 7

8

9

Page 8: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

8

0

1 2

3

4 5

6 7

8

9

This neural network could manage to

correctly recognise about 9 out of every 10 digits

Deep-­‐learning  methods  are  representa0on-­‐learning  methods  with  mul>ple  levels  of  

representa>on,  obtained  by  composing  simple  but  non-­‐linear  modules  that  each  transform  the  representa>on  at  one  level  (star>ng  with  the  raw  input)  into  a  representa>on  at  a  higher,  slightly  

more  abstract  level.    [LeCun  etal,  2014]  

Deep Learning Yann LeCun, Yoshua Bengio & Geoffrey Hinton http://www.nature.com/nature/journal/v521/n7553/full/nature14539.html

Page 9: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

9

0 1 2 3 4 5 6 7 8

9

0 1 2 3 4 5 6 7 8

9

A deep learning appraoch could

manage to correctly recognise about 10

out of every 10 digits

Page 10: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

10

Deep  neural  networks  seem  to  brilliantly  address  the  selec0vity-­‐invariance  dilemma  that  is  fundamental  to  all  efforts  to  learn  to  classify  objects:  they  produce  representa>ons  that  are  selec>ve  to    the  aspects  of  the  image  that  are  important  for  discrimina>on,  but  that  are  

invariant  to  irrelevant  aspects    

Deep  networks  hold  records  for  problems  in  image  recogni0on,  speech  recogni0on,  and  text  

classifica0on  amongst  other  areas  

Hardware  Data  Algorithms  

Applica>ons  

Page 11: Brian Mac Namee - Predict Webinar 3 - Short Intro to Deep Learing

13/08/15

11

Thank You

Questions? Fundamentals of Machine Learning for Predictive Data Analytics: Algorithms, Worked Examples, and Case Studies John D. Kelleher, Brian Mac Namee and Aoife D'Arcy www.machinelearningbook.com