目前做法整理

$W_{mean} = \sum_{in} W_{in,out}/N_{in}$
$W^{'}= W - W_{mean}$
$fakeout = X*W^{'}$
$W^{''} = W^{'}/\sigma(fakeout)$
$loss_{weight} = ||X*W^{''} - X*W^{'}||_{2}$
$realout = X*W^{''}$
$Loss_{total}=loss_{classfication}+\lambda *loss_{weight}$