<s>
Labeled	B-General_Concept
data	I-General_Concept
is	O
a	O
group	O
of	O
samples	O
that	O
have	O
been	O
tagged	O
with	O
one	O
or	O
more	O
labels	O
.	O
</s>
<s>
Labeled	B-General_Concept
data	I-General_Concept
is	O
significantly	O
more	O
expensive	O
to	O
obtain	O
than	O
the	O
raw	O
unlabeled	O
data	O
.	O
</s>
<s>
In	O
2006	O
Fei-Fei	O
Li	O
,	O
the	O
co-director	O
of	O
the	O
Stanford	O
Human-Centered	O
AI	B-Application
Institute	O
,	O
set	O
out	O
to	O
improve	O
the	B-Application
artificial	I-Application
intelligence	I-Application
models	O
and	O
algorithms	O
for	O
image	O
recognition	O
by	O
significantly	O
enlarging	O
the	O
training	O
data	O
.	O
</s>
<s>
The	O
3.2	O
million	O
images	O
that	O
were	O
labelled	O
by	O
more	O
than	O
49,000	O
workers	O
formed	O
the	O
basis	O
for	O
ImageNet	B-General_Concept
,	O
one	O
of	O
the	O
largest	O
hand-labeled	O
database	O
for	O
outline	O
of	O
object	O
recognition	O
.	O
</s>
<s>
Training	O
data	O
that	O
relies	O
on	O
bias	O
labeled	B-General_Concept
data	I-General_Concept
will	O
result	O
in	O
prejudices	O
and	O
omissions	O
in	O
a	O
predictive	B-General_Concept
model	I-General_Concept
,	O
despite	O
the	O
machine	O
learning	O
algorithm	O
being	O
legitimate	O
.	O
</s>
<s>
Because	O
the	O
labeled	B-General_Concept
data	I-General_Concept
available	O
to	O
train	O
facial	O
recognition	O
systems	O
has	O
not	O
been	O
representative	O
of	O
a	O
population	O
,	O
underrepresented	O
groups	O
in	O
the	O
labeled	B-General_Concept
data	I-General_Concept
are	O
later	O
often	O
misclassified	O
.	O
</s>
