Never Ending Learning Mitchell video
<p>Oct 29, 2012 11:02 PM
 
Tom Mitchell  CMU
 
humans learn many things for years  and become better learners over time...
ride bikes,  tennis,  pick up stuff... we get better at it over decades...
 
most programs live a very short life...  they get runed off...
 
why not machines?  cumulative...  underexplored
 
Never Ending learning
 
Task  acquire a growing compeence without asymptote
over years, mult funciton
where learning one thinks improves its ability etc.
 
lecture at UW  Seattle...
 
robots how to not run out of power...
how not to run into things...
robots, softbots...
 
this was 2010 lecture...
 
architectures for problem solving  SOAR  rosenblood
Icasus  langley  and Prodigy  carbonnell
 
large scale kno cosntruct  cyc, knowitall, textrunner, woe
 
life long learing  
learning to learn  thrun and pratt 98  and thrun and mitchel 93
 
transfer learning  multitask  caruana  95
transfer reinforcment leaning parr and russell 98
learning with struc outpus  roth 2009  taskar 2009
 
active larning  survey  multitask leaning
 
curric learning  bengio  2009
 
transfer learning... coupling... eg  spam filter1  helps u learn spam filter 2...
 
Cast study we're trying to work on
 
NELL  Never-Ending Language Learner...
 
input an ontology
 
eg  give it a dozen univ examples...  occas intearction with aperson...
 
each day it reads more facts to populate ontology  and it needs to
do it better(eg  more facts more accurately out of the text)
 
turned on in Jan 2010  running 24 by 7
 
500 catgories and raltions
10 to 20 see examples
500M  examples
 
since Jan  executing an algo...
 
now has 400K  extracted triples...
 
 
look at 
 
READ THE WEB... look at what its learning... its all online
 
eg gazebo is a kind of furniture
 
alex barzuk is
 
jimi is a musician who plays the guitar...
 
melbourne museum
 
its learning triples...
 
he looks at seattle...
 
is a city...
 
and provides evidence... v high confidence...
looking at half a billion web pages...  collecting
 
it cooccurs with other words  that indicates its a city... using bayes...
 
incorrect inferences do pollute the DB and v quickly..
 
is seattle  the capit of wash......
 
has newspaper stranger;  mariners, seahawks 
 
gray stuff... does not quite believe...
confience below some threshold...
 
 
400,000 of these triples... can browse at your leisure...
 
worked on this for 3 years  floundering around
 
2 good ideas  that make the algo...
 
its classifying noun phrases... do they pertain to a city...
 
NELL has 20 examples that are labelled...
and has 20M  unlabeled...
 
semi-supervised bootstrap learning...
 
eg  given cities  like pitssue  paris, seatteee, cupterino
mayor of...
live in
 
live in state of denial
 
but  it is underconstrained...  a bad example  gets in there  (like live in a stae of denial)
 
guess what  vague concept i'm thing of...
 
key idea 1:  its too difficult... to learn  an unconstrained  thing  but
 
force it to learn a lot of things
 
couplelearning things together...
 
3 difft types of coupling of functions...
 
minute 21/22
coupled training tpe 1: 
 
co-training....
 
type 1  coupling constraints in NELL...
 
cotraining...  multiple labesl  provide a joint constraint...
 
 
type 2  use same x  going to difft Ys...
 
X  goes to  Y1  is it a city  ot
Y2 is it a fruit;
 
is it a fruit  (then its not a city)  and vice versa...
 
 
athlete implies person
 
athlete implies not sport
 
sport +> not athlete...
 
we input by hand  some constraints...
 
third constraint...
 
athlete plays for team   or athlete plays sport...
 
what are the tpes of the arguemtns...
 
Alex Rodriguez   plays for NY Yankees  then it must also say that  A rod is an athlete...
 
its simult learning a 1000 difft  relations...
key idea number 1  is above..
 
*
algo is  like EM
Pure EM appraoch to coupled training..
 
E  jointly estimate latent labels for each func of each unlabeled example
 
M  retrain all functions
 
locally max  likelihood...
 
but does work
 
there really are  1014 NP paris...  like to do it  but  can't do it.. too expensive..
 
NELL  does E'M'   
E' stepjust does this for the ones that its the most confident of
top 250  NPs... next iteration... does next 250...
 
M' step: 
 
Q does it retract things that it prev thot was true... NO... but it should
 
NELL has a KB.... grwoing set of beleifs that its decided to commit to...
 
evidence integrator...   looks at candidate beleifs...
 
simple  iterative loop...
 
what set of beleifs should we beleive...  simple strategy...
 
arg1 is the tiget woods of arg2...
 
eg  pikes  peak  or   last word  is mountain
 
mountain bike is not a mountain...
 
the bdeatles  is a good morphology
 
does it end in sun for being a newspaper...
 
obelinsky... prob. a name...
 
morpholog  analysis  give remarkably good resuslts
 
silliam cohen and richard wang  SEAL...  book author
 
3 types of learned extractors...  they make indep errors
if error is 20%  then combined error is only .23... good error rate...
 
Coupled Training Helps...
 
Precisions really improve  up into 90% with coupled constraints...
 
food is  100%  movie is 100%
 
 but sports eqpt  is terrible...
 
maybe  need a new predicate...
 
eg add human traists  (for denial)
 
coupled traiing  is a good idea...
 
so how can we get new constraints...
 
how abt getting NELL to LEARN the constraints...
 
key idea 2:  DISCOVER New Coupling Constraints...
 
first order, proab horn clause constraints...
 
eg  learn this rule
 
if  A rod plays for Yankees then infer that  A rod plays baseball...
 
here's a bad rule it learned
 
UW  plays basketball therefore  UW plays in the NBA
 
had not read about college basketball...
 
 
comment from Dan:  nowehre in the web will u find a statemtn that 
UW basketball team does NOT playin the national hockey league;
 
learned probab  rules  ... horn clause rules...
 
NELL as of Oct 18 2010...
 
 
in June  humans got into the loop  June 2010
 
after 10 iterations,... say  forget all these  emotions... take abt 5 mins every 2 weeks
to correct it...
 
 
Q is it getting better at learning... is it happening...
 
its happening for majoring of predicates but getting worse for a mnority...
 
baked good  ran amok  (cuz of internet cookie problem)
 
some new stuff...
 
NELL has a fixed ontology... not useful to be fixed...
 
 
Mahamed and Hruschka   similar to Oren's and my work...
 
 
consider musician  and musical instrument...
 
eg jimi hendriks  and guitar...
 
Arg 1 master  Arg 2
 
sitar  George Harrison
tenor Sax  Stan Getz...
 
manually adding  some relations...
 
 
type has member.... eg  gas  has member  co2...
 
ontology extending by reading...
 
can u popularity and discusison on web to extend it
 
Wonderful game that will be on the
 
POLARITY  play this game
 
Edith Law and Burr Settles and Luis von Ahn...  
 
do this....   crowd sourcing... keep NELL on track...
 
 
bottom line:
 
1)  think abt how  wide open  never ending learning is...nobody is working on this...
lots of great phd theses....
turn it into a never ending learning robot
how never ending learning street lamp  turn green or red or call police
never ending  chess player...
 
 
 
 
 </p>