<s>
In	O
machine	O
learning	O
,	O
boosting	B-Algorithm
is	O
an	O
ensemble	B-Algorithm
meta-algorithm	B-Algorithm
for	O
primarily	O
reducing	O
bias	O
,	O
and	O
also	O
variance	O
in	O
supervised	B-General_Concept
learning	I-General_Concept
,	O
and	O
a	O
family	O
of	O
machine	O
learning	O
algorithms	O
that	O
convert	O
weak	B-Algorithm
learners	I-Algorithm
to	O
strong	O
ones	O
.	O
</s>
<s>
Boosting	B-Algorithm
is	O
based	O
on	O
the	O
question	O
posed	O
by	O
Kearns	O
and	O
Valiant	O
(	O
1988	O
,	O
1989	O
)	O
:	O
"	O
Can	O
a	O
set	O
of	O
weak	B-Algorithm
learners	I-Algorithm
create	O
a	O
single	O
strong	O
learner	O
?	O
"	O
</s>
<s>
A	O
weak	B-Algorithm
learner	I-Algorithm
is	O
defined	O
to	O
be	O
a	O
classifier	B-General_Concept
that	O
is	O
only	O
slightly	O
correlated	O
with	O
the	O
true	O
classification	O
(	O
it	O
can	O
label	O
examples	O
better	O
than	O
random	O
guessing	O
)	O
.	O
</s>
<s>
In	O
contrast	O
,	O
a	O
strong	O
learner	O
is	O
a	O
classifier	B-General_Concept
that	O
is	O
arbitrarily	O
well-correlated	O
with	O
the	O
true	O
classification	O
.	O
</s>
<s>
Robert	O
Schapire	O
's	O
affirmative	O
answer	O
in	O
a	O
1990	O
paper	O
to	O
the	O
question	O
of	O
Kearns	O
and	O
Valiant	O
has	O
had	O
significant	O
ramifications	O
in	O
machine	O
learning	O
and	O
statistics	O
,	O
most	O
notably	O
leading	O
to	O
the	O
development	O
of	O
boosting	B-Algorithm
.	O
</s>
<s>
When	O
first	O
introduced	O
,	O
the	O
hypothesis	O
boosting	B-Algorithm
problem	O
simply	O
referred	O
to	O
the	O
process	O
of	O
turning	O
a	O
weak	B-Algorithm
learner	I-Algorithm
into	O
a	O
strong	O
learner	O
.	O
</s>
<s>
"	O
Informally	O
,	O
[	O
the	O
hypothesis	O
boosting ]	O
problem	O
asks	O
whether	O
an	O
efficient	O
learning	O
algorithm	O
 [ … ] 	O
that	O
outputs	O
a	O
hypothesis	O
whose	O
performance	O
is	O
only	O
slightly	O
better	O
than	O
random	O
guessing	O
[	O
i.e.	O
</s>
<s>
Algorithms	O
that	O
achieve	O
hypothesis	O
boosting	B-Algorithm
quickly	O
became	O
simply	O
known	O
as	O
"	O
boosting	B-Algorithm
"	O
.	O
</s>
<s>
Freund	O
and	O
Schapire	O
's	O
arcing	O
(	O
Adapt[at]ive	O
Resampling	O
and	O
Combining	O
)	O
,	O
as	O
a	O
general	O
technique	O
,	O
is	O
more	O
or	O
less	O
synonymous	O
with	O
boosting	B-Algorithm
.	O
</s>
<s>
While	O
boosting	B-Algorithm
is	O
not	O
algorithmically	O
constrained	O
,	O
most	O
boosting	B-Algorithm
algorithms	O
consist	O
of	O
iteratively	O
learning	O
weak	B-Algorithm
classifiers	I-Algorithm
with	O
respect	O
to	O
a	O
distribution	O
and	O
adding	O
them	O
to	O
a	O
final	O
strong	O
classifier	B-General_Concept
.	O
</s>
<s>
When	O
they	O
are	O
added	O
,	O
they	O
are	O
weighted	O
in	O
a	O
way	O
that	O
is	O
related	O
to	O
the	O
weak	B-Algorithm
learners	I-Algorithm
 '	O
accuracy	O
.	O
</s>
<s>
After	O
a	O
weak	B-Algorithm
learner	I-Algorithm
is	O
added	O
,	O
the	O
data	O
weights	O
are	O
readjusted	O
,	O
known	O
as	O
"	O
re-weighting	O
"	O
.	O
</s>
<s>
Thus	O
,	O
future	O
weak	B-Algorithm
learners	I-Algorithm
focus	O
more	O
on	O
the	O
examples	O
that	O
previous	O
weak	B-Algorithm
learners	I-Algorithm
misclassified	O
.	O
</s>
<s>
There	O
are	O
many	O
boosting	B-Algorithm
algorithms	O
.	O
</s>
<s>
The	O
original	O
ones	O
,	O
proposed	O
by	O
Robert	O
Schapire	O
(	O
a	O
recursive	O
majority	O
gate	O
formulation	O
)	O
and	O
Yoav	O
Freund	O
(	O
boost	O
by	O
majority	O
)	O
,	O
were	O
not	O
adaptive	B-Algorithm
and	O
could	O
not	O
take	O
full	O
advantage	O
of	O
the	O
weak	B-Algorithm
learners	I-Algorithm
.	O
</s>
<s>
Schapire	O
and	O
Freund	O
then	O
developed	O
AdaBoost	B-Algorithm
,	O
an	O
adaptive	B-Algorithm
boosting	B-Algorithm
algorithm	O
that	O
won	O
the	O
prestigious	O
Gödel	O
Prize	O
.	O
</s>
<s>
Only	O
algorithms	O
that	O
are	O
provable	O
boosting	B-Algorithm
algorithms	O
in	O
the	O
probably	O
approximately	O
correct	O
learning	O
formulation	O
can	O
accurately	O
be	O
called	O
boosting	B-Algorithm
algorithms	O
.	O
</s>
<s>
Other	O
algorithms	O
that	O
are	O
similar	O
in	O
spirit	O
to	O
boosting	B-Algorithm
algorithms	O
are	O
sometimes	O
called	O
"	O
leveraging	O
algorithms	O
"	O
,	O
although	O
they	O
are	O
also	O
sometimes	O
incorrectly	O
called	O
boosting	B-Algorithm
algorithms	O
.	O
</s>
<s>
The	O
main	O
variation	O
between	O
many	O
boosting	B-Algorithm
algorithms	O
is	O
their	O
method	O
of	O
weighting	B-General_Concept
training	O
data	O
points	O
and	O
hypotheses	O
.	O
</s>
<s>
AdaBoost	B-Algorithm
is	O
very	O
popular	O
and	O
the	O
most	O
significant	O
historically	O
as	O
it	O
was	O
the	O
first	O
algorithm	O
that	O
could	O
adapt	O
to	O
the	O
weak	B-Algorithm
learners	I-Algorithm
.	O
</s>
<s>
It	O
is	O
often	O
the	O
basis	O
of	O
introductory	O
coverage	O
of	O
boosting	B-Algorithm
in	O
university	O
machine	O
learning	O
courses	O
.	O
</s>
<s>
There	O
are	O
many	O
more	O
recent	O
algorithms	O
such	O
as	O
LPBoost	B-Algorithm
,	O
TotalBoost	O
,	O
BrownBoost	B-Algorithm
,	O
xgboost	B-Language
,	O
MadaBoost	O
,	O
LogitBoost	B-Algorithm
,	O
and	O
others	O
.	O
</s>
<s>
Many	O
boosting	B-Algorithm
algorithms	O
fit	O
into	O
the	O
AnyBoost	O
framework	O
,	O
which	O
shows	O
that	O
boosting	B-Algorithm
performs	O
gradient	B-Algorithm
descent	I-Algorithm
in	O
a	O
function	B-Algorithm
space	I-Algorithm
using	O
a	O
convex	O
cost	B-Algorithm
function	I-Algorithm
.	O
</s>
<s>
Given	O
images	O
containing	O
various	O
known	O
objects	O
in	O
the	O
world	O
,	O
a	O
classifier	B-General_Concept
can	O
be	O
learned	O
from	O
them	O
to	O
automatically	O
classify	B-General_Concept
the	O
objects	O
in	O
future	O
images	O
.	O
</s>
<s>
Simple	O
classifiers	B-General_Concept
built	O
based	O
on	O
some	O
image	B-Algorithm
feature	I-Algorithm
of	O
the	O
object	O
tend	O
to	O
be	O
weak	O
in	O
categorization	O
performance	O
.	O
</s>
<s>
Using	O
boosting	B-Algorithm
methods	O
for	O
object	B-General_Concept
categorization	I-General_Concept
is	O
a	O
way	O
to	O
unify	O
the	O
weak	B-Algorithm
classifiers	I-Algorithm
in	O
a	O
special	O
way	O
to	O
boost	O
the	O
overall	O
ability	O
of	O
categorization	O
.	O
</s>
<s>
Object	B-General_Concept
categorization	I-General_Concept
is	O
a	O
typical	O
task	O
of	O
computer	B-Application
vision	I-Application
that	O
involves	O
determining	O
whether	O
or	O
not	O
an	O
image	O
contains	O
some	O
specific	O
category	O
of	O
object	O
.	O
</s>
<s>
Appearance	O
based	O
object	B-General_Concept
categorization	I-General_Concept
typically	O
contains	O
feature	B-Algorithm
extraction	I-Algorithm
,	O
learning	O
a	O
classifier	B-General_Concept
,	O
and	O
applying	O
the	O
classifier	B-General_Concept
to	O
new	O
examples	O
.	O
</s>
<s>
from	O
shape	B-Algorithm
analysis	I-Algorithm
,	O
bag	B-General_Concept
of	I-General_Concept
words	I-General_Concept
models	O
,	O
or	O
local	O
descriptors	O
such	O
as	O
SIFT	B-Algorithm
,	O
etc	O
.	O
</s>
<s>
Examples	O
of	O
supervised	B-General_Concept
classifiers	I-General_Concept
are	O
Naive	B-General_Concept
Bayes	I-General_Concept
classifiers	I-General_Concept
,	O
support	B-Algorithm
vector	I-Algorithm
machines	I-Algorithm
,	O
mixtures	O
of	O
Gaussians	O
,	O
and	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
However	O
,	O
research	O
has	O
shown	O
that	O
object	O
categories	O
and	O
their	O
locations	O
in	O
images	O
can	O
be	O
discovered	O
in	O
an	O
unsupervised	B-General_Concept
manner	I-General_Concept
as	O
well	O
.	O
</s>
<s>
The	O
recognition	O
of	O
object	O
categories	O
in	O
images	O
is	O
a	O
challenging	O
problem	O
in	O
computer	B-Application
vision	I-Application
,	O
especially	O
when	O
the	O
number	O
of	O
categories	O
is	O
large	O
.	O
</s>
<s>
This	O
is	O
due	O
to	O
high	O
intra	O
class	O
variability	O
and	O
the	O
need	O
for	O
generalization	B-Algorithm
across	O
variations	O
of	O
objects	O
within	O
the	O
same	O
category	O
.	O
</s>
<s>
Even	O
the	O
same	O
object	O
may	O
appear	O
unalike	O
under	O
different	O
viewpoint	O
,	O
scale	B-Algorithm
,	O
and	O
illumination	O
.	O
</s>
<s>
One	O
means	O
is	O
by	O
feature	B-Algorithm
sharing	O
and	O
boosting	B-Algorithm
.	O
</s>
<s>
AdaBoost	B-Algorithm
can	O
be	O
used	O
for	O
face	O
detection	O
as	O
an	O
example	O
of	O
binary	B-General_Concept
categorization	I-General_Concept
.	O
</s>
<s>
After	O
boosting	B-Algorithm
,	O
a	O
classifier	B-General_Concept
constructed	O
from	O
200	O
features	O
could	O
yield	O
a	O
95%	O
detection	O
rate	O
under	O
a	O
false	O
positive	O
rate	O
.	O
</s>
<s>
Another	O
application	O
of	O
boosting	B-Algorithm
for	O
binary	B-General_Concept
categorization	I-General_Concept
is	O
a	O
system	O
that	O
detects	O
pedestrians	O
using	O
patterns	O
of	O
motion	O
and	O
appearance	O
.	O
</s>
<s>
It	O
takes	O
a	O
similar	O
approach	O
to	O
the	O
Viola-Jones	B-General_Concept
object	I-General_Concept
detection	I-General_Concept
framework	I-General_Concept
.	O
</s>
<s>
Compared	O
with	O
binary	B-General_Concept
categorization	I-General_Concept
,	O
multi-class	B-General_Concept
categorization	I-General_Concept
looks	O
for	O
common	O
features	O
that	O
can	O
be	O
shared	O
across	O
the	O
categories	O
at	O
the	O
same	O
time	O
.	O
</s>
<s>
They	O
turn	O
to	O
be	O
more	O
generic	O
edge	B-Algorithm
like	O
features	O
.	O
</s>
<s>
Compared	O
with	O
training	O
separately	O
,	O
it	O
generalizes	B-Algorithm
better	O
,	O
needs	O
less	O
training	O
data	O
,	O
and	O
requires	O
fewer	O
features	O
to	O
achieve	O
the	O
same	O
performance	O
.	O
</s>
<s>
During	O
each	O
iteration	O
the	O
algorithm	O
chooses	O
a	O
classifier	B-General_Concept
of	O
a	O
single	O
feature	B-Algorithm
(	O
features	O
that	O
can	O
be	O
shared	O
by	O
more	O
categories	O
shall	O
be	O
encouraged	O
)	O
.	O
</s>
<s>
This	O
can	O
be	O
done	O
via	O
converting	O
multi-class	B-General_Concept
classification	I-General_Concept
into	O
a	O
binary	O
one	O
(	O
a	O
set	O
of	O
categories	O
versus	O
the	O
rest	O
)	O
,	O
or	O
by	O
introducing	O
a	O
penalty	O
error	O
from	O
the	O
categories	O
that	O
do	O
not	O
have	O
the	O
feature	B-Algorithm
of	O
the	O
classifier	B-General_Concept
.	O
</s>
<s>
In	O
the	O
paper	O
"	O
Sharing	O
visual	B-Algorithm
features	I-Algorithm
for	O
multiclass	O
and	O
multiview	O
object	O
detection	O
"	O
,	O
A	O
.	O
Torralba	O
et	O
al	O
.	O
</s>
<s>
used	O
GentleBoost	O
for	O
boosting	B-Algorithm
and	O
showed	O
that	O
when	O
training	O
data	O
is	O
limited	O
,	O
learning	O
via	O
sharing	O
features	O
does	O
a	O
much	O
better	O
job	O
than	O
no	O
sharing	O
,	O
given	O
same	O
boosting	B-Algorithm
rounds	O
.	O
</s>
<s>
Also	O
,	O
for	O
a	O
given	O
performance	O
level	O
,	O
the	O
total	O
number	O
of	O
features	O
required	O
(	O
and	O
therefore	O
the	O
run	O
time	O
cost	O
of	O
the	O
classifier	B-General_Concept
)	O
for	O
the	O
feature	B-Algorithm
sharing	O
detectors	O
,	O
is	O
observed	O
to	O
scale	B-Algorithm
approximately	O
logarithmically	O
with	O
the	O
number	O
of	O
class	O
,	O
i.e.	O
,	O
slower	O
than	O
linear	O
growth	O
in	O
the	O
non-sharing	O
case	O
.	O
</s>
<s>
Similar	O
results	O
are	O
shown	O
in	O
the	O
paper	O
"	O
Incremental	O
learning	O
of	O
object	O
detectors	O
using	O
a	O
visual	O
shape	O
alphabet	O
"	O
,	O
yet	O
the	O
authors	O
used	O
AdaBoost	B-Algorithm
for	O
boosting	B-Algorithm
.	O
</s>
<s>
Boosting	B-Algorithm
algorithms	O
can	O
be	O
based	O
on	O
convex	O
or	O
non-convex	O
optimization	O
algorithms	O
.	O
</s>
<s>
Convex	O
algorithms	O
,	O
such	O
as	O
AdaBoost	B-Algorithm
and	O
LogitBoost	B-Algorithm
,	O
can	O
be	O
"	O
defeated	O
"	O
by	O
random	O
noise	O
such	O
that	O
they	O
ca	O
n't	O
learn	O
basic	O
and	O
learnable	O
combinations	O
of	O
weak	O
hypotheses	O
.	O
</s>
<s>
However	O
,	O
by	O
2009	O
,	O
multiple	O
authors	O
demonstrated	O
that	O
boosting	B-Algorithm
algorithms	O
based	O
on	O
non-convex	O
optimization	O
,	O
such	O
as	O
BrownBoost	B-Algorithm
,	O
can	O
learn	O
from	O
noisy	O
datasets	O
and	O
can	O
specifically	O
learn	O
the	O
underlying	O
classifier	B-General_Concept
of	O
the	O
Long	O
–	O
Servedio	O
dataset	O
.	O
</s>
