<s>
Visual	B-Application
temporal	I-Application
attention	I-Application
is	O
a	O
special	O
case	O
of	O
visual	O
attention	O
that	O
involves	O
directing	O
attention	O
to	O
specific	O
instant	O
of	O
time	O
.	O
</s>
<s>
Similar	O
to	O
its	O
spatial	O
counterpart	O
visual	O
spatial	O
attention	O
,	O
these	O
attention	O
modules	O
have	O
been	O
widely	O
implemented	O
in	O
video	O
analytics	O
in	O
computer	B-Application
vision	I-Application
to	O
provide	O
enhanced	O
performance	O
and	O
human	O
interpretable	O
explanation	O
of	O
deep	B-Algorithm
learning	I-Algorithm
models	O
.	O
</s>
<s>
As	O
visual	O
spatial	O
attention	O
mechanism	O
allows	O
human	O
and/or	O
computer	B-Application
vision	I-Application
systems	O
to	O
focus	O
more	O
on	O
semantically	O
more	O
substantial	O
regions	O
in	O
space	O
,	O
visual	B-Application
temporal	I-Application
attention	I-Application
modules	O
enable	O
machine	O
learning	O
algorithms	O
to	O
emphasize	O
more	O
on	O
critical	O
video	O
frames	O
in	O
video	O
analytics	O
tasks	O
,	O
such	O
as	O
human	B-Application
action	I-Application
recognition	I-Application
.	O
</s>
<s>
Research	O
in	O
human	B-Application
action	I-Application
recognition	I-Application
has	O
accelerated	O
significantly	O
since	O
the	O
introduction	O
of	O
powerful	O
tools	O
such	O
as	O
Convolutional	B-Architecture
Neural	I-Architecture
Networks	I-Architecture
(	O
CNNs	B-Architecture
)	O
.	O
</s>
<s>
However	O
,	O
effective	O
methods	O
for	O
incorporation	O
of	O
temporal	O
information	O
into	O
CNNs	B-Architecture
are	O
still	O
being	O
actively	O
explored	O
.	O
</s>
<s>
Motivated	O
by	O
the	O
popular	O
recurrent	O
attention	O
models	O
in	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
the	O
Attention-aware	O
Temporal	O
Weighted	O
CNN	B-Architecture
(	O
ATW	O
CNN	B-Architecture
)	O
is	O
proposed	O
in	O
videos	O
,	O
which	O
embeds	O
a	O
visual	O
attention	O
model	O
into	O
a	O
temporal	O
weighted	O
multi-stream	O
CNN	B-Architecture
.	O
</s>
<s>
Besides	O
,	O
each	O
stream	O
in	O
the	O
proposed	O
ATW	O
CNN	B-Architecture
framework	O
is	O
capable	O
of	O
end-to-end	O
training	O
,	O
with	O
both	O
network	O
parameters	O
and	O
temporal	O
weights	O
optimized	O
by	O
stochastic	B-Algorithm
gradient	I-Algorithm
descent	I-Algorithm
(	O
SGD	O
)	O
with	O
back-propagation	B-Algorithm
.	O
</s>
<s>
Experimental	O
results	O
show	O
that	O
the	O
ATW	O
CNN	B-Architecture
attention	O
mechanism	O
contributes	O
substantially	O
to	O
the	O
performance	O
gains	O
with	O
the	O
more	O
discriminative	O
snippets	O
by	O
focusing	O
on	O
more	O
relevant	O
video	O
segments	O
.	O
</s>
