Machine learning

Post on 16-Nov-2014

408 views 3 download

description

 

Transcript of Machine learning

Machine Learning Overview

Ricardo Wendell

June 2013

2

Artificial Intelligence & Machine Learning Artificial Neural Networks Clustering Genetic Algorithms Reinforcement Learning Q&A

Agenda

3

Artificial Intelligence & Machine Learning

What is Intelligence?

4

What’s involved in intelligence?

ü Ability to interact with the real world

ü Reasoning and planning

ü Learning and adaptation

5

What’s involved in intelligence?

ü Ability to interact with the real world

ü Reasoning and planning

ü Learning and adaptation

6

“To learn is to gain knowledge, or

understanding of, or skill in, by study, instruction, or experience”

7

Machine Learning

Field of study that gives computers the ability to learn without being explicitly programmed

8

Why?

9

Why?

Why should machines have to learn?

10

Why?

Why should machines have to learn? Why not design machines to perform as desired in the first place?

11

Some reasons…

There are tasks that cannot be defined well except by example

It is possible that hidden among large piles of data are important relationships and correlations

12

An example

13 http://googleresearch.blogspot.com.br/2012/08/speech-recognition-and-deep-learning.html

Google’s Deep Learning

How can machines learn?

14

Types of learning

Supervised

Types of learning

Unsupervised

16

Supervised

Ways of address these problems…

ü Statistics

ü Brain Models

ü Adaptive Control

ü Evolutionary Models

ü Psychological Models

17

Ways of address these problems…

ü Statistics

ü Brain Models

ü Adaptive Control

ü Evolutionary Models

ü Psychological Models

18

19

Artificial Neural Networks

Biological Inspiration

20

McCulloch-Pitts Neuron (1943)

21

Multilayer Perceptron

22

How do we train it?

23

Backpropagation Algorithm initialize  the  weights  in  the  network  

do  

   for  each  example  e  in  the  training  set  

       O  =  neural-­‐net-­‐output(network,  i)  

       T  =  teacher  output  for  e  

       compute  error  (T  -­‐  O)  at  the  output  units  

       compute  delta_wh  for  all  weights  from  hidden  layer  to  output  layer  

       compute  delta_wi  for  all  weights  from  input  layer  to  hidden  layer  

       update  the  weights  in  the  network  

until  all  examples  classified  correctly  or  stopping  criterion  satisfied  

return  the  network  

24

Recurrent ANN

25

Drawbacks

ANNs with many hidden layers can solve difficult problems… but are very hard to train!

26

Deep Learning http://www.kurzweilai.net/how-bio-inspired-deep-learning-keeps-winning-competitions 27

28

Clustering

Clustering algorithms

29

Clustering problems

http://jsfiddle.net/8NpNp/27/ 30

Many applications

Medical imaging

Market research

Climatology

Social networks analysis

Recommender systems

31

32

Genetic Algorithms

Given a population, how can we find the best

group of individuals that satisfy some criteria?

33

Given a population, how can we find the best

group of individuals that satisfy some criteria?

34

Typical requirements

a genetic representation of the solution domain

a fitness function to evaluate the solution domain

35

Genetic representation

Each solution is represented as a “chromosome”

Each solution has a fitness value

36

Generic Genetic Algorithm

37

Example: Mona Lisa from 1500 characters

38 http://www.youtube.com/watch?v=TManzvC9pi8&NR

39

Reinforcement Learning

Main concepts

Inspired by behaviorist psychology

Learning by interacting with the environment

40

Suited for problems which include a long-term versus short-term reward trade-off

How to teach a computer to play games?

41

Markov Decision Processes

42

How to flip pancakes?

43

http://www.youtube.com/watch?v=W_gxLKSsSIE

44

Conclusion

Conclusion There are many techniques…

Results are heavily influenced by input representation

It’s a math-heavy field!

45

Some libraries

46

http://www.youtube.com/watch?v=WB9zr0IZCPQ

47

Q&A