Introduc)on

 to  Machine  
Learning:  CS  436/580L  
Introduc)on  

Instructor:  Ar)  Ramesh  

Binghamton  University  
 Logis)cs  
•  Instructor:  Ar)  Ramesh  
–  Email:  ar)r@binghamton.edu  
–  Office:  N-­‐4  Engineering  Building  
–  Office  hours:  Tuesdays/Thursdays  –  2:45  p.m.  to  3:45  p.m.  
•  TA:  Yue  Zhang  
–  Email  yzhan202@binghamton.edu    
–  Office  hours:  4  -­‐  6  pm  Friday,  or  by  appointment  
•  Course  available  now  on  myCourses  

2
Grading  
•  Five  homeworks  (30%)    
–  6%  each  
–  Due  one-­‐two  weeks  later  
–  Some  programming,  some  exercises  
–  Assigned  via  myCourses.  
•  One  Midterm  (20%),  One  Final  (25%)  
–  Exams  are  closed  book.  You  will  be  allowed  a  cheat  
sheet,  a  double-­‐sided  8.5  x  11  page.  
•  Quizzes  (5%)  

3
Grading  
•  Project  (15%)  
–  Groups  of  three  
•  Class  Par)cipa)on  (5%)  
–  based  on  abendance  and  how  ac)vely  a  student  
par)cipates  in  class  discussions.  
•  Abendance  and  par)cipa)on  is  mandatory  
–  Grade  reduced  by  a  leber  grade  for  lack  of  abendance  
(e.g.,  A-­‐  becomes  B-­‐;  B-­‐  becomes  a  C-­‐;  etc)  
–  Miss  three  classes  without  prior  approval  will  receive  an  
F  in  the  class  

4
Project  
•  Class  project  info  –  Get  into  teams  of  size  3  as  
soon  as  you  can  
–  Group  forming  deadline:  Feb  5th,  2017  
–  Would  prefer  size  3  groups,  but  I  might  allow  some  
groups  of  size  2  if  there  is  a  strong  reason  
–  individual  projects  are  strongly  discouraged  
–  If  you  are  unable  to  form  a  group,  email  me/TAs  
and  we  will  assign  you  to  a  group.  

5
Project  Pre-­‐proposal  
•  Pre-­‐proposal  due  on  Feb  5th  
–  One  person  from  each  team  email  me  and  the  TAs  
the  team  members  along  with  an  abstract  
–  Should  include  topic,  possible  datasets/algorithms  
you  plan  to  explore  

6
Source  Materials  
•  T.  Mitchell,  Machine  Learning,  
McGraw-­‐Hill    (required)  
•  C.  Bishop,  PaRern  Recogni)on  and  Machine  
Learning,  Springer  (required)  
•  Kevin  Murphy,  Machine  Learning:  A  
probabilis)c  perspec)ve  (recommended)  
•  Class  Notes/Slides  

7
Class  Ac)vity  
•  Discuss  with  your  neighbor  
•  What  do  you  think  is  machine  learning?  Define  
in  your  own  words.  
•  Think  of  one  task  you  think  that  is  already  using/
could  greatly  benefit  from  machine  learning.    
 E.g.,  self-­‐driving  cars!  
Be  crea)ve!!!  I  would  like  as  many  different  
answers  as  possible!!!  
 
8
Introducing  Machine  Learning  

Play  video:  
hbps://www.ted.com/talks/
jeremy_howard_the_wonderful_and_terrifying_i
mplica)ons_of_computers_that_can_learn#t-­‐484
56  

9
 Glassdoor:  Data  Scien)st  Jobs    
Best Jobs

25 Best Jobs in America 7.6k

Shares

Want a new job? Glassdoor is here to help, identifying the 25 Best Jobs in America for 2016. The jobs that make
this list have the highest overall Glassdoor Job Score, determined by combining three key factors – number of job
openings, salary and career opportunities rating. These jobs stand out across all three categories.

 
United States 2016

1 Data Scientist
Job Openings 1,736
Median Base Salary $116,840
Career Opportunity 4.1
Job Score 4.7

2 Tax Manager
Job Openings 1,574
Median Base Salary $108,000
10
 http://www.bloomberg.com/news/articles/
2015-06-04/help-wanted-black-belts-in-data

11
 What  is  Data  Science?  

12
 Most  Important  Technical  Skills  for  
Data  Science  

13
 Most  Important  Technical  Skills  for  
Data  Science  

14
 Why  Study  Machine  Learning:  
A  Few  Quotes  
•  “A  breakthrough  in  machine  learning  would  be  worth  
ten  Microsons”  (Bill  Gates,  Microson)  
•  “Machine  learning  is  the  next  Internet”    
(Tony  Tether,  Former  Director,  DARPA)  
•  Machine  learning  is  the  hot  new  thing”    
(John  Hennessy,  President,  Stanford)  
•  “Web  rankings  today  are  mostly  a  maber  of  machine  
learning”  (Prabhakar  Raghavan,  Former  Dir.  Research,  
Yahoo)  
•  “Machine  learning  is  going  to  result  in  a  real  
revolu)on”  (Greg  Papadopoulos,  CTO,  Sun)  
 
15
 •  Geqng  computers  to  program  
themselves  
•  Wri)ng  sonware  is  the  
bobleneck,  let  data  do  the  work  
   Tradi)onal  Programming  
Requirements   Program   Output  
Data   Human   Computer  

Input  
Requirements  and  data  change  o^en  

   Machine  Learning  
Requirements   Output  
Machine   Program  
Data   Computer  
Learning  
Input  

16
 Training  Data  
Training  Example  

Two    
Classes:  
{Yes,No}  

17
 •  Geqng  computers  to  program  
themselves  
•  Wri)ng  sonware  is  the  
bobleneck,  let  data  do  the  work  
   Tradi)onal  Programming  
Data  
Computer   Output  
Program  

   Machine  Learning  
Data  
Computer   Program  
Output  
18
 Magic?    
No,  more  like  gardening  
 
•  Seeds  =  Algorithms  
•  Nutrients  =  Data  
•  Gardener  =  You  
•  Plants  =  Programs  

19
 Defini)on:  Machine  Learning!  
•  T.  Mitchell:  Well  posed  machine  learning  
–  Improving  performance  via  experience  
–  Formally,  A  computer  program  is  said  to  learn  from  
experience  E  with  respect  to  some  class  of  tasks  T  and  
performance  measure  P,  it  its  performance  at  tasks  in  T  as  
measured  by  P,  improves  with  experience.  
•  H.  Simon  
–  Learning  denotes  changes  in  the  system  that  are  adap)ve  in  
the  sense  that  they  enable  the  system  to  do  the  task  or  tasks  
drawn  from  the  same  popula)on  more  efficiently  and  more  
effec)vely  the  next  )me.  
 The  ability  to  perform  a  task  in  a  situa)on  which  has  never  been  
encountered  before  

20
 Example  1:  A  Chess  learning  problem  
•  Task  T:  playing  chess  
•  Performance  measure  P:  percent  of  games  won  
against  opponents  
•  Training  Experience  E:  playing  prac)ce  games  
against  itself  

21
 Example  2:  Autonomous  Vehicle  
Problem  
•  Task  T:  driving  on  a  public  highway/roads  using  
vision  sensors  
•  Performance  Measure  P:  percentage  of  )me  the  
vehicle  is  involved  in  an  accident  
•  Training  Experience  E:    a  sequence  of  images  
and  steering  commands  recorded  while  
observing  a  human  driver  

22
 When  to  use  Machine  Learning?  
•  Human  exper)se  is  absent  
–  Example:  naviga)ng  on  mars  
•  Humans  are  unable  to  explain  their  
exper)se  
–  Example:  vision,  speech,  language  
•  Requirements  and  data  change  over  
)me  
–  Example:  Tracking,  Biometrics,  
Personalized  fingerprint  recogni)on  
•  The  problem  or  the  data  size  is  just  
too  large  
–  Example:  Web  Search  

23
 Types  of  Learning  
•  Supervised  (induc)ve)  learning  
–  Training  data  includes  desired  outputs  
•  Unsupervised  learning  
–  Training  data  does  not  include  desired  outputs  
–  Find  hidden/interes)ng  structure  in  data  
•  Semi-­‐supervised  learning  
–  Training  data  includes  a  few  desired  outputs  
•  Reinforcement  learning  
–  the  learner  interacts  with  the  world  via  “ac)ons”  and  
tries  to  find  an  op)mal  policy  of  behavior  with  respect  
to  “rewards”  it  receives  from  the  environment  

24
 Examples/Types  of  Machine  Learning  
Tasks  
•  Forecas)ng  or  Predic)on  
–  Stock  price  of  Google  tomorrow?  
•  Classifica)on  and  Regression  
–  Is  Ana  credit-­‐worthy?  
–  What  is  Ana’s  credit  score?  
•  Ranking  
–  How  to  rank  images  that  contain  “An  awesome  machine  learning  
model”?  
•  Outlier/Anomaly/Fraud  detec)on  
–  Is  it  Ana”  using  the  credit  card  in  Mexico  or  is  it  someone  else?  
•  Finding  paberns  
–  Almost  60%  of  shoppers  buy  Diapers  and  Milk  together!  

25
 Machine  Learning:  Applica)ons  
•  Examples  of  real-­‐world  scenarios  where  
machine  learning  performs  its  magic!  

26
 Classifica)on  Example:  Spam  Filtering  

Classify  as  “Spam”  or  “Not  Spam”  

27
 Classifica)on  Example:  Weather  
Predic)on  

28
 Regression  example:  Predic)ng  Gold/
Stock  prices  

Good  ML  can  make  

you  rich  (but  there  
is  s)ll  some  risk  
involved).  

Given  historical  data  on  Gold  prices,  

predict  tomorrow’s  price!  
29
 Similarity  Determina)on  

30
 Collabora)ve  Filtering  
•  The  problem  of  collabora/ve  filtering  is  to  
predict  how  well  a  user  will  like  an  item  that  he  
has  not  rated  given  a  set  of  historical  preference  
judgments  for  a  community  of  users.  
 

31
 Collabora)ve  Filtering  

32
 Collabora)ve  Filtering  

33
 Clustering:  Discover  Structure  in  data  

34
 Clustering  

35
 ML  in  a  Nutshell  
•  Tens  of  thousands  of  machine  learning  
algorithms  
•  Hundreds  new  every  year  
•  Every  machine  learning  algorithm  has  three  
components:  
–  Representa)on  
–  Evalua)on  
–  Op)miza)on  

36
 Representa)on  
•  Decision  trees  
•  Sets  of  rules  /  Logic  programs  
•  Instances  
•  Graphical  models  (Bayes/Markov  nets)  
•  Neural  networks  
•  Support  vector  machines  
•  Model  ensembles  
•  Etc.  

37
 Evalua)on  
•  Accuracy  
•  Precision  and  recall  
•  Squared  error  
•  Likelihood  
•  Posterior  probability  
•  Cost  /  U)lity  
•  Margin  
•  Entropy  
•  K-­‐L  divergence  
•  Etc.  

38
 Op)miza)on  
•  Combinatorial  op)miza)on  
–  E.g.:  Greedy  search  
•  Convex  op)miza)on  
–  E.g.:  Gradient  descent  
•  Constrained  op)miza)on  
–  E.g.:  Linear  programming  

39
 Machine  learning  has  grown  in  leaps  
and  bounds  
•  The  main  approach  for  
–  Speech  Recogni)on  
–  Robo)cs  
–  Natural  Language  Processing  
–  Computa)onal  Biology  
–  Sensor  networks  
–  Computer  Vision  
–  Web  
–  And  so  on  
40
 What  We’ll  Cover  
•  Supervised  learning:  Decision  tree  induc)on,  Rule  induc)on,  
Instance-­‐based  learning,  Bayesian  learning,  Neural  networks,  
Support  vector  machines,  Linear  Regression,  Model  ensembles  
•  Unsupervised  learning:  Clustering,  Dimensionality  reduc)on  
•  General  machine  learning  concepts  and  techniques:  
Feature  selec)on,  cross-­‐valida)on,  maximum  likelihood  
es)ma)on,  gradient  descent,  expecta)on-­‐maximiza)on  
•  And  some  special  topics  (if  )me  permits):  probabilis)c  
graphical  models,  topic  models  
•  Your  responsibility:  
–  Brush  up  on  some  important  background  
–  Linear  algebra,  Sta)s)cs  101,  Vectors,  Probability  theory  
 

41