<s>
Active	B-General_Concept
learning	I-General_Concept
is	O
a	O
special	O
case	O
of	O
machine	O
learning	O
in	O
which	O
a	O
learning	O
algorithm	O
can	O
interactively	O
query	O
a	O
user	O
(	O
or	O
some	O
other	O
information	O
source	O
)	O
to	O
label	O
new	O
data	O
points	O
with	O
the	O
desired	O
outputs	O
.	O
</s>
<s>
This	O
type	O
of	O
iterative	O
supervised	O
learning	O
is	O
called	O
active	B-General_Concept
learning	I-General_Concept
.	O
</s>
<s>
Recent	O
developments	O
are	O
dedicated	O
to	O
multi-label	O
active	B-General_Concept
learning	I-General_Concept
,	O
hybrid	O
active	B-General_Concept
learning	I-General_Concept
and	O
active	B-General_Concept
learning	I-General_Concept
in	O
a	O
single-pass	O
(	O
on-line	O
)	O
context	O
,	O
combining	O
concepts	O
from	O
the	O
field	O
of	O
machine	O
learning	O
(	O
e.g.	O
</s>
<s>
conflict	O
and	O
ignorance	O
)	O
with	O
adaptive	O
,	O
incremental	B-Algorithm
learning	I-Algorithm
policies	O
in	O
the	O
field	O
of	O
online	B-Algorithm
machine	I-Algorithm
learning	I-Algorithm
.	O
</s>
<s>
Large-scale	O
active	B-General_Concept
learning	I-General_Concept
projects	O
may	O
benefit	O
from	O
crowdsourcing	O
frameworks	O
such	O
as	O
Amazon	O
Mechanical	O
Turk	O
that	O
include	O
many	O
humans	O
in	O
the	O
active	B-General_Concept
learning	I-General_Concept
loop	O
.	O
</s>
<s>
Most	O
of	O
the	O
current	O
research	O
in	O
active	B-General_Concept
learning	I-General_Concept
involves	O
the	O
best	O
method	O
to	O
choose	O
the	O
data	O
points	O
for	O
.	O
</s>
<s>
This	O
strategy	O
manages	O
this	O
compromise	O
by	O
modelling	O
the	O
active	B-General_Concept
learning	I-General_Concept
problem	O
as	O
a	O
contextual	O
bandit	O
problem	O
.	O
</s>
<s>
Expected	O
error	O
reduction	O
:	O
label	O
those	O
points	O
that	O
would	O
most	O
reduce	O
the	O
model	O
's	O
generalization	B-Algorithm
error	I-Algorithm
.	O
</s>
<s>
Exponentiated	O
Gradient	O
Exploration	O
for	O
Active	B-General_Concept
Learning	I-General_Concept
:	O
In	O
this	O
paper	O
,	O
the	O
author	O
proposes	O
a	O
sequential	O
algorithm	O
named	O
exponentiated	O
gradient	O
(	O
EG	O
)	O
-active	O
that	O
can	O
improve	O
any	O
active	B-General_Concept
learning	I-General_Concept
algorithm	O
by	O
an	O
optimal	O
random	O
exploration	O
.	O
</s>
<s>
Margin	B-Algorithm
Sampling	O
:	O
The	O
sample	O
with	O
the	O
smallest	O
difference	O
between	O
the	O
two	O
highest	O
class	O
probabilities	O
is	O
considered	O
to	O
be	O
the	O
most	O
uncertain	O
.	O
</s>
<s>
Querying	O
from	O
diverse	O
subspaces	O
or	O
partitions	O
:	O
When	O
the	O
underlying	O
model	O
is	O
a	O
forest	O
of	O
trees	O
,	O
the	O
leaf	O
nodes	O
might	O
represent	O
(	O
overlapping	O
)	O
partitions	O
of	O
the	O
original	O
feature	B-Algorithm
space	I-Algorithm
.	O
</s>
<s>
Conformal	O
predictors	O
:	O
This	B-Algorithm
method	I-Algorithm
predicts	O
that	O
a	O
new	O
data	O
point	O
will	O
have	O
a	O
label	O
similar	O
to	O
old	O
data	O
points	O
in	O
some	O
specified	O
way	O
and	O
degree	O
of	O
the	O
similarity	O
within	O
the	O
old	O
examples	O
is	O
used	O
to	O
estimate	O
the	O
confidence	O
in	O
the	O
prediction	O
.	O
</s>
<s>
Some	O
active	B-General_Concept
learning	I-General_Concept
algorithms	O
are	O
built	O
upon	O
support-vector	B-Algorithm
machines	I-Algorithm
(	O
SVMs	B-Algorithm
)	O
and	O
exploit	O
the	O
structure	O
of	O
the	O
SVM	B-Algorithm
to	O
determine	O
which	O
data	O
points	O
to	O
label	O
.	O
</s>
<s>
Such	O
methods	O
usually	O
calculate	O
the	O
margin	B-Algorithm
,	O
,	O
of	O
each	O
unlabeled	O
datum	O
in	O
and	O
treat	O
as	O
an	O
-dimensional	O
distance	O
from	O
that	O
datum	O
to	O
the	O
separating	O
hyperplane	O
.	O
</s>
<s>
Minimum	O
Marginal	O
Hyperplane	O
methods	O
assume	O
that	O
the	O
data	O
with	O
the	O
smallest	O
are	O
those	O
that	O
the	O
SVM	B-Algorithm
is	O
most	O
uncertain	O
about	O
and	O
therefore	O
should	O
be	O
placed	O
in	O
to	O
be	O
labeled	O
.	O
</s>
