<s>
Multi-task	B-General_Concept
learning	I-General_Concept
(	O
MTL	O
)	O
is	O
a	O
subfield	O
of	O
machine	O
learning	O
in	O
which	O
multiple	O
learning	O
tasks	O
are	O
solved	O
at	O
the	O
same	O
time	O
,	O
while	O
exploiting	O
commonalities	O
and	O
differences	O
across	O
tasks	O
.	O
</s>
<s>
In	O
a	O
widely	O
cited	O
1997	O
paper	O
,	O
Rich	O
Caruana	O
gave	O
the	O
following	O
characterization:Multitask	O
Learning	O
is	O
an	O
approach	O
to	O
inductive	B-General_Concept
transfer	I-General_Concept
that	O
improves	O
generalization	B-Algorithm
by	O
using	O
the	O
domain	O
information	O
contained	O
in	O
the	O
training	O
signals	O
of	O
related	O
tasks	O
as	O
an	O
inductive	B-General_Concept
bias	I-General_Concept
.	O
</s>
<s>
It	O
does	O
this	O
by	O
learning	O
tasks	O
in	O
parallel	O
while	O
using	O
a	O
shared	O
representation	B-General_Concept
;	O
what	O
is	O
learned	O
for	O
each	O
task	O
can	O
help	O
other	O
tasks	O
be	O
learned	O
better	O
.	O
</s>
<s>
Further	O
examples	O
of	O
settings	O
for	O
MTL	O
include	O
multiclass	B-General_Concept
classification	I-General_Concept
and	O
multi-label	B-Algorithm
classification	I-Algorithm
.	O
</s>
<s>
Multi-task	B-General_Concept
learning	I-General_Concept
works	O
because	O
regularization	O
induced	O
by	O
requiring	O
an	O
algorithm	O
to	O
perform	O
well	O
on	O
a	O
related	O
task	O
can	O
be	O
superior	O
to	O
regularization	O
that	O
prevents	O
overfitting	B-Error_Name
by	O
penalizing	O
all	O
complexity	O
uniformly	O
.	O
</s>
<s>
For	O
example	O
,	O
with	O
sparsity	B-Algorithm
,	O
overlap	O
of	O
nonzero	O
coefficients	O
across	O
tasks	O
indicates	O
commonality	O
.	O
</s>
<s>
Novel	O
methods	O
which	O
builds	O
on	O
a	O
prior	O
multitask	O
methodology	O
by	O
favoring	O
a	O
shared	O
low-dimensional	O
representation	B-General_Concept
within	O
each	O
task	O
grouping	O
have	O
been	O
proposed	O
.	O
</s>
<s>
Experiments	O
on	O
synthetic	O
and	O
real	O
data	O
have	O
indicated	O
that	O
incorporating	O
unrelated	O
tasks	O
can	O
result	O
in	O
significant	O
improvements	O
over	O
standard	O
multi-task	B-General_Concept
learning	I-General_Concept
methods	O
.	O
</s>
<s>
Related	O
to	O
multi-task	B-General_Concept
learning	I-General_Concept
is	O
the	O
concept	O
of	O
knowledge	O
transfer	O
.	O
</s>
<s>
Whereas	O
traditional	O
multi-task	B-General_Concept
learning	I-General_Concept
implies	O
that	O
a	O
shared	O
representation	B-General_Concept
is	O
developed	O
concurrently	O
across	O
tasks	O
,	O
transfer	O
of	O
knowledge	O
implies	O
a	O
sequentially	O
shared	O
representation	B-General_Concept
.	O
</s>
<s>
Large	O
scale	O
machine	O
learning	O
projects	O
such	O
as	O
the	O
deep	B-Architecture
convolutional	I-Architecture
neural	I-Architecture
network	I-Architecture
GoogLeNet	O
,	O
an	O
image-based	O
object	O
classifier	B-General_Concept
,	O
can	O
develop	O
robust	O
representations	O
which	O
may	O
be	O
useful	O
to	O
further	O
algorithms	O
learning	O
related	O
tasks	O
.	O
</s>
<s>
Traditionally	O
Multi-task	B-General_Concept
learning	I-General_Concept
and	O
transfer	O
of	O
knowledge	O
are	O
applied	O
to	O
stationary	O
learning	O
settings	O
.	O
</s>
<s>
The	O
form	O
of	O
the	O
kernel	O
induces	O
both	O
the	O
representation	B-General_Concept
of	O
the	O
feature	O
space	O
and	O
structures	O
the	O
output	O
across	O
tasks	O
.	O
</s>
<s>
This	O
factorization	O
property	O
,	O
separability	O
,	O
implies	O
the	O
input	O
feature	O
space	O
representation	B-General_Concept
does	O
not	O
vary	O
by	O
task	O
.	O
</s>
<s>
For	O
the	O
separable	O
case	O
,	O
the	O
representation	B-General_Concept
theorem	O
is	O
reduced	O
to	O
.	O
</s>
<s>
Letting	O
,	O
where	O
is	O
the	O
Laplacian	B-Algorithm
for	O
the	O
graph	O
with	O
adjacency	B-Algorithm
matrix	I-Algorithm
M	O
giving	O
pairwise	O
similarities	O
of	O
tasks	O
.	O
</s>
<s>
Non-convex	O
penalties	O
-	O
Penalties	O
can	O
be	O
constructed	O
such	O
that	O
A	B-Application
is	I-Application
constrained	O
to	O
be	O
a	O
graph	B-Algorithm
Laplacian	I-Algorithm
,	O
or	O
that	O
A	O
has	O
low	O
rank	O
factorization	O
.	O
</s>
<s>
In	O
large	O
scale	O
open	O
membership	O
email	O
systems	O
,	O
most	O
users	O
do	O
not	O
label	O
enough	O
messages	O
for	O
an	O
individual	O
local	O
classifier	B-General_Concept
to	O
be	O
effective	O
,	O
while	O
the	O
data	O
is	O
too	O
noisy	O
to	O
be	O
used	O
for	O
a	O
global	O
filter	O
across	O
all	O
users	O
.	O
</s>
<s>
A	O
hybrid	O
global/individual	O
classifier	B-General_Concept
can	O
be	O
effective	O
at	O
absorbing	O
the	O
influence	O
of	O
users	O
who	O
label	O
emails	O
very	O
diligently	O
from	O
the	O
general	O
public	O
.	O
</s>
<s>
Using	O
boosted	O
decision	B-Algorithm
trees	I-Algorithm
,	O
one	O
can	O
enable	O
implicit	O
data	O
sharing	O
and	O
regularization	O
.	O
</s>
<s>
Here	O
,	O
multitask	B-General_Concept
learning	I-General_Concept
is	O
particularly	O
helpful	O
as	O
data	O
sets	O
from	O
different	O
countries	O
vary	O
largely	O
in	O
size	O
because	O
of	O
the	O
cost	O
of	O
editorial	O
judgments	O
.	O
</s>
<s>
The	O
Multi-Task	B-General_Concept
Learning	I-General_Concept
via	O
StructurAl	O
Regularization	O
(	O
MALSAR	O
)	O
Matlab	O
package	O
implements	O
the	O
following	O
multi-task	B-General_Concept
learning	I-General_Concept
algorithms	O
:	O
</s>
